spec : fix n-gram drafts rejected at temp > 0 after truncation

[praneshgo]

Overview

Fix the regression where, at temperature > 0, n-gram drafts cut short near the end of a reply were wrongly rejected outright, slowing generation down. The regression affects n-gram drafters and is fixed with this PR; draft-simple and draft-mtp are not affected by the regression.
Regression introduced by #27694 commit 1fb7ef3e3.

Root cause

• The server passes a candidate list (result_q) to the drafters whenever temperature > 0. use_spec_rejection() checks only temp > 0.0f, and --spec-draft-sampling doesn't affect this.

• draft-simple and draft-mtp either clear the list in greedy mode or fill one candidate list per drafted token in probabilistic mode. N-gram drafters (and EAGLE3/DFlash/DSpark) never touch it.

• common_speculative_draft() truncates drafts longer than dp.n_max, which is the remaining room before max_tokens or the end of the context. It also resizes result_q to n_max. On an untouched, empty list this creates n_max empty candidate lists.

• The verifier then sees a non-empty list whose size matches the draft, so it switches to rejection sampling. With q = 0 for every drafted token, it rejects the whole draft.

• Effect: near the end of a reply, every truncated n-gram draft is fully rejected.

• Both greedy (default) and probabilistic modes are affected, temperature 0 is not affected, draft-simple and draft-mtp on their own are not affected.

Fix

• When a draft is truncated, the candidate list is now trimmed only if it isn't empty, i.e. only if the drafter actually produced candidates.
• An empty list stays empty, so n-gram drafts are verified by greedy sample-and-match, as before #27694.
• Drafters that do produce candidates (draft-simple and draft-mtp in probabilistic mode) still have their list trimmed together with the draft, as #27694 intended.

Testing

Tested on RTX 5090, following decode perf data is with builds corresponding to pre-PR 134b2bb, ToT 836d571, ToT + fix (this PR). Temperature 0.7 over 8 seeds:

Llama-3.1-8B-Instruct Q4_K_M (dense):

--spec-type	build	median t/s	worst seed t/s	mean drafted	mean acceptance
ngram-mod	pre-PR	509.6	419.1	1048	0.493
	ToT	375.1	334.7	1368	0.365
	fix	500.7	419.0	1048	0.493

Qwen3.6-35B-A3B Q4_K_M (hybrid, built-in MTP):

--spec-type	build	median t/s	worst seed t/s	mean drafted	mean acceptance
ngram-mod,draft-mtp (np=1)	ToT	301.7	203.1	1306	0.496
	fix	333.9	304.9	1168	0.553
ngram-mod	ToT	227.8	184.9	849	0.377
	fix	292.8	262.0	780	0.514
ngram-mod,draft-mtp (np=2)	ToT	187.0	145.9	1254	0.538
	fix	329.4	225.4	1041	0.670
ngram-mod,draft-mtp (--spec-draft-sampling probabilistic)	ToT	307.6	218.4	1233	0.522
	fix	344.1	314.2	1214	0.529

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[praneshgo]

@ggerganov, @danbev, @gaugarg-nv, @lbyte
Can you please review this PR? Thank you.

[lbyte]

Verified, the regression is fixed.
Thanks!