Clear the textmap cache in Page.flush_cache()

[siddharthgaur1]

Fixes #1395.

Problem

Page.flush_cache() iterates Page.cached_properties, which covers _layout and the object/edge caches — but not the lru_cache wrapping _get_textmap, which is created as an instance attribute in Page.__init__ and is the largest of a page's caches. So a flush left most of the memory in place.

Measured on tests/pdfs/chelsea_pdta.pdf (65 pages), calling extract_text() on every page under tracemalloc:

	traced memory
no flush	184.61 MB
flush_cache() — before this PR	105.19 MB
close()	4.80 MB
flush_cache() — after this PR	4.80 MB

Note on the issue's diagnosis

The report attributes the retention to _layout surviving the flush and to page_obj, and proposes page.page_obj = None as the fix. Neither holds up:

• _layout is cleared already. flush_cache reads self.cached_properties, which on a Page resolves to Container.cached_properties + ["_layout"], so the default flush includes it.
• page.page_obj = None changes nothing measurable (105.19 MB → 105.19 MB). The entire difference is the textmap cache.

The retention the reporter measured is real; the cause is the textmap cache.

Change

• Page.flush_cache() now clears the textmap cache after delegating to Container.flush_cache.
• Page.close() did this explicitly already, so it now just delegates to flush_cache().
• DerivedPage.__init__ builds its textmap cache before calling flush_cache() rather than after, since flush_cache() now touches that attribute. Without the reorder, constructing any cropped page raises AttributeError.

Users who want the old surgical behaviour can still pass an explicit list, e.g. page.flush_cache(["_objects"]) — the textmap cache is cleared either way, which matches what the method's name promises.

Tests

Adds test_flush_cache_clears_textmap, covering both an original page and a cropped one (the DerivedPage path that the reorder protects). It fails on the parent commit (assert 1 == 0) and passes with the change.

Full suite: 164 passed, 10 failed — all 10 failures are pre-existing on a clean develop in this environment (test_repair.py needs Ghostscript, test_convert.py hits Windows line-ending differences) and are unrelated to this change.

black, isort --profile black, flake8 and mypy all pass on the touched files.