docs: shorten the batching mode title and rework the flownode diagrams

[killme2008]

What changed

Rename the Flownode batching mode page to Batching Mode. The old heading
Flownode Batching Mode Developer Guide also drove the sidebar label, which was
too long for the flownode section. Cross-references in dataflow.md and
overview.md follow the new name.

Replace batching_mode_arch.png with an inline mermaid diagram, matching the
plugin added in #2801 and the existing usage in enterprise/iceberg.md. The
rendered diagram goes from 633x1582 to 387x608, and mermaid picks up the site's
configured light and dark themes instead of shipping a light-only raster.

The old image also had errors, which the new diagram fixes:

• It drew BatchingEngine and Database as decision diamonds. Neither is a decision.
• Its Notify edge implied that a source-table write drives the task. docs: correct outdated Nightly contributor guide content #2784
  corrected this: marking a window dirty does not wake the task, which runs on its
  evaluation schedule or adaptive polling cadence.
• Database is vague. The task encodes an INSERT INTO sink SELECT plan and sends
  it through FrontendClient, so the node is now Frontend.
• Frontend scans the source table while executing that plan. The old image had no
  source-table read edge.

static/batching_mode_arch.png is kept rather than deleted, in case the URL is
referenced outside this repository.

Flow mode selection

dataflow.md listed the mode-selection rules without their precedence, so a reader
could conclude that an aggregating Flow on a source table with ttl = 'instant' runs
in batching mode. determine_flow_type checks the source tables first: a missing
source table needs defer_on_missing_source and yields a pending batching Flow, an
instant-TTL source forces streaming, and only then are TQL and Aggregate/Distinct
considered. The paragraph is replaced with a flowchart in that order, plus an explicit
sentence about the instant-TTL consequence.

v1.0 gets a shorter diagram. defer_on_missing_source does not exist in v1.0.2, where
an instant-TTL source returns streaming directly from the source-table loop.

Scope

• Documentation versions: Nightly, 1.2, 1.1, 1.0
• Languages: English, Chinese

Verification

• DOC_LANG=en pnpm check:links
• DOC_LANG=zh pnpm check:links
• git diff --check
• Rendered every built page in headless Chrome for all four versions and both
  locales: the H1 is Batching Mode / 批处理模式, and each page emits exactly one
  aria-roledescription="flowchart" SVG.
• Verified the mermaid blocks are byte-identical across the versions that share them,
  and that the v1.0 flow-selection diagram omits the defer_on_missing_source branch.
• Verified the diagram against GreptimeDB main: gen_insert_plan_unlocked builds a
  LogicalPlan::Dml(WriteOp::Insert) over the sink table, encode_insert_plan_request
  wraps it as api::v1::InsertIntoPlan, and it is sent through FrontendClient, so
  Frontend performs both the source scan and the sink write.
• No lockfile or package.json change.

Checklist

• I verified the content against the applicable GreptimeDB version.
• I updated the relevant documentation versions and languages, or explained why not.
• I checked changed links and anchors.
• I updated navigation when the document structure changed. No sidebar file changes
  are needed: the entry is the doc id, and its label comes from the H1.

[cloudflare-workers-and-pages]

Deploying greptime-docs with    Cloudflare Pages

Latest commit:	5748d9a
Status:	✅  Deploy successful!
Preview URL:	https://409bffdd.greptime-docs.pages.dev
Branch Preview URL:	https://docs-shorten-batching-mode-t.greptime-docs.pages.dev

View logs

[discord9]

The reported Flow execution-path and v1.0 behavior issues are fixed. I also verified that the renamed H1s do not generate Docusaurus fragment IDs, so no AnchorAlias is needed. The release-specific EN/ZH content is aligned, and all checks pass.