fix: support page:last doclib locators

[lylelllll]

Motivation

mineru read "doc:<short_id>/tier:flash/page:last" is currently rejected even when the document's last page is cached. Numeric page locators work, but callers cannot use the documented last-page shorthand.

Fixes #5300

Modification

• Accept last as a page token in read locators.
• Resolve page:last against the document's page_count before the existing numeric page-range and target logic runs.
• Return canonical numeric locators and content ranges in the response.
• Return a clear invalid_locator error when the document page count is unavailable.
• Add regression coverage for both the successful read path and the unavailable-page-count path.

BC-breaking

No. Existing numeric page, block, and character locators retain their current behavior and remain 1-based.

Testing

• .venv/bin/python -m pytest -q -o addopts='' tests/unittest/test_doclib_cache_semantics.py tests/unittest/test_doclib_locators.py tests/unittest/test_doclib_progressive_reading.py (210 passed)
• uvx ruff check mineru/doclib/server.py
• uvx ruff check --ignore ANN001,E501 tests/unittest/test_doclib_cache_semantics.py (the ignored rules have existing violations outside this patch)
• Direct parser/resolution smoke test for page:last/block:2/char:8

Model/GPU parsing E2E was not run because this change is isolated to doclib locator parsing and cached-content reads.

Checklist

• Linting was run on the modified production code.
• The reported bug is covered by regression tests.
• Existing numeric locator and progressive-reading tests pass.
• No downstream API or cache-format change is introduced.
• CLA check passed.

[myhloli]

Thank you for the contribution. Last-page selection is now supported on dev in aca9b06da through the same syntax used by parsing inputs.

Use page:r1 for the last page and page:r2 for the second-to-last page. Locators also accept ranges, comma-separated selections and all, such as page:r3-r1 and page:1,3,r1. Returned references use absolute page numbers. Relative selections require a known document page count; block/character and image reads require a single resolved page.

Range reads preserve the selected pages during continuation and stop when the selection is complete. Local validation: 667 related Doclib/CLI tests passed, with Ruff and formatting checks passing.

The page:last spelling proposed here is superseded by the shared page-selection syntax, so I am closing this PR.

Refs #5300.