Add optional segment_boxes to /markdown and /html to skip re-analysis

[makothap]

Motivation

Callers that already ran layout analysis on a document — most commonly via
POST / — have no way to reuse those segments when subsequently calling
/markdown or /html. Today both endpoints always re-run analysis
internally, even when the caller already has an equivalent, freshly-computed
segment list in hand. For a pipeline that calls POST / once and then
/markdown and/or /html for output, that's up to 3 analysis passes over the
same document where 1 would do.

This mirrors a pattern already in this codebase: /toc_from_xml accepts a
segment_boxes form field and skips analysis entirely, using the caller's
segments directly. This PR extends the same idea to the two conversion
endpoints.

Design

• segment_boxes: optional JSON-encoded list of segment dicts (same shape as
  POST /'s response), added as a Form field on /markdown and /html,
  parsed the same way /toc_from_xml already parses its own segment_boxes.
• In ConvertToMarkdownUseCase.execute / ConvertToHtmlUseCase.execute, when
  segment_boxes is provided it's used as the analysis result directly
  (skipping both analyze_pdf_layout and analyze_pdf_layout_fast); when
  absent, behavior is completely unchanged (including fast mode routing).
• No changes to PDFAnalysisService, the conversion services, or the
  Clean-Architecture layering — segment_boxes flows in as a use-case
  parameter, not a driver-level hack, consistent with extract_toc,
  dpi, etc.
• Malformed JSON in segment_boxes raises JSONDecodeError, already caught
  by the existing @catch_exceptions decorator and turned into a 422, so
  no new error handling was needed.

Evidence

From pepng (a downstream consumer, not part of this PR) reprocessing the same
fixtures with and without segment reuse — end-to-end pipeline duration,
INGEST started → COMPLETE completed:

fixture	3 passes (no reuse)	1 pass (this feature)	reduction
table-bordered.pdf	6.975s	2.277s	−67.3%
table-merged.pdf	6.540s	2.205s	−66.3%
twocol-continuous.pdf	7.808s	3.142s	−59.8%

The engine's own logs confirm the skip (this PR adds the equivalent log
lines, matching /toc_from_xml's existing logging style):

[INFO] Calling endpoint: convert_to_html_endpoint
[INFO] Received 3 segment boxes for /html; skipping analysis.

Output was verified byte-for-byte identical with vs. without segment reuse
across three fixtures (document.md, document.html, document.json).

Test coverage

• Unit tests (src/use_cases/markdown_conversion/test_convert_to_markdown_use_case.py,
  src/use_cases/html_conversion/test_convert_to_html_use_case.py, new in
  this PR): analysis is skipped when segment_boxes is given (asserted via a
  PDFAnalysisService stub that raises if called), and unchanged (including
  fast-mode routing) when it isn't.
• End-to-end tests (src/tests/test_end_to_end.py, new in this PR, run
  against the real service):
  • /markdown and /html output is identical with vs. without reusing
    segment_boxes from a prior POST / call, on test_pdfs/regular.pdf.
  • Malformed segment_boxes ("not-json") returns 422 for both
    endpoints, not 500.
• Full existing suite run against the patched build, all 34 tests
  passing (src/tests/test_end_to_end.py, includes the 4 new tests above):
  built natively on a DGX Spark (ARM64, NVIDIA GB10 GPU) via
  docker build --build-arg BUILDER_IMAGE=nvidia/cuda:13.0.3-cudnn-devel-ubuntu24.04 --build-arg TORCH_INDEX_URL=https://download.pytorch.org/whl/cu130 --build-arg TORCH_CUDA_ARCH_LIST=12.0;12.1,
  run on a non-default port, verified against the live GPU-backed engine (not
  mocked), then torn down.
• black --check --line-length 125 clean on all changed files.
• Not run: the project's CI only targets ubuntu-latest/amd64; this was
  verified on aarch64 instead as that's the available hardware. No GPU-less
  CPU-only run was attempted (out of scope — this change has no CPU/GPU
  branching).

What wasn't run and why

• The repo's own CI (test.yml, GitHub Actions, amd64) wasn't triggered —
  this is a fork branch, not yet a PR, so Actions haven't run. All tests
  above were run manually against a natively-built image on aarch64
  hardware instead.
• Translation-related tests (Ollama-backed) weren't specifically re-verified
  for this change since segment_boxes doesn't interact with translation;
  the existing translation tests are part of the "full suite" pass above and
  passed unmodified.

[ali6parmak]

Thanks for the PR, merging.