[Feature]Support tag phase token enforce generation

[freeliuzc]

Motivation

1. 确保模型生成 \n\n\n 后，生成 <tool_call> / 其中一个token
   [Feature]Support tag phase token enforce generation #6031

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

Usage or Command

Accuracy Tests

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

函数缺少docstring说明。作为新增的公共API，应该添加详细的文档字符串，包括：

1. 函数的功能描述
2. 各个参数的含义和预期类型
3. 返回值说明
4. 使用示例（如果适用）
   这对于API的可维护性和可用性非常重要。

Suggested change

	reasoning_phase_token_constraint
	Apply reasoning-phase token constraints to the logits for next-token prediction.
	This function wraps a device-specific backend operator that enforces constraints
	on the token distribution during the reasoning (or "thinking") phase. It updates
	``logits`` in place according to the current decoding step, allowed tokens, and
	reasoning status, then returns the same tensor for convenience.
	Args:
	logits (paddle.Tensor): Logits for candidate tokens at the current decoding
	step. Shape is typically ``[batch_size, vocab_size]`` or a similar
	2-D tensor. This tensor is modified in place.
	pre_token_ids (paddle.Tensor): Previously generated token ids for each
	sequence in the batch. Typically a 2-D integer tensor of shape
	``[batch_size, seq_len]``.
	stop_flags (paddle.Tensor): Boolean or integer tensor indicating which
	sequences have already met their stop conditions and should no longer
	produce new tokens. Shape: ``[batch_size]``.
	seq_lens_this_time (paddle.Tensor): Effective sequence lengths for the
	current decoding step (after padding is removed). Shape: ``[batch_size]``.
	seq_lens_encoder (paddle.Tensor): Encoder-side sequence lengths associated
	with each sequence in the batch. Shape: ``[batch_size]``.
	step_idx (paddle.Tensor): Current decoding step index, usually a scalar or
	a 1-D tensor with a single element.
	reasoning_allowed_tokens (paddle.Tensor): Tensor containing the ids of
	tokens that are allowed to be generated during the reasoning phase.
	The exact shape depends on how allowed tokens are represented but is
	typically integer-typed.
	reasoning_status (paddle.Tensor): Tensor indicating the reasoning status
	for each sequence (for example, whether the sequence is currently in
	reasoning mode or in normal generation). Shape: ``[batch_size]``.
	output_padding_offset (paddle.Tensor): Tensor used to map from padded
	batch positions to compacted positions after removing padding.
	output_cum_offsets (paddle.Tensor): Cumulative offsets associated with
	``output_padding_offset`` for efficient indexing on the device.
	think_end_id (int): Special token id that marks the end of the reasoning
	phase. When this id is generated, constraints may change or be lifted.
	line_break_id (int): Token id corresponding to a line break (such as
	``"\\n"``). This can be used as a constraint boundary for reasoning
	or output formatting.
	Returns:
	paddle.Tensor: The same ``logits`` tensor after in-place modification by
	the backend operator. Returned for API convenience and potential chaining.
	Example:
	.. code-block:: python
	logits = reasoning_phase_token_constraint(
	logits=logits,
	pre_token_ids=pre_token_ids,
	stop_flags=stop_flags,
	seq_lens_this_time=seq_lens_this_time,
	seq_lens_encoder=seq_lens_encoder,
	step_idx=step_idx,
	reasoning_allowed_tokens=reasoning_allowed_tokens,
	reasoning_status=reasoning_status,
	output_padding_offset=output_padding_offset,
	output_cum_offsets=output_cum_offsets,
	think_end_id=think_end_id,
	line_break_id=line_break_id,
	)

[codecov-commenter]

Codecov Report

❌ Patch coverage is 53.84615% with 6 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@2c17acd). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...or/layers/sample/ops/apply_penalty_multi_scores.py	20.00%	4 Missing ⚠️
fastdeploy/model_executor/layers/sample/sampler.py	60.00%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6034   +/-   ##
==========================================
  Coverage           ?   67.33%           
==========================================
  Files              ?      349           
  Lines              ?    45069           
  Branches           ?     6939           
==========================================
  Hits               ?    30345           
  Misses             ?    12486           
  Partials           ?     2238           

Flag	Coverage Δ
GPU	67.33% <53.84%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.