feat(api): expose strategy auto-selection on the public API

[jerefrer]

Closes #18.

Strategy auto-selection currently lives only in cli_main, so the patch_pdf / extract_pdf_text / build_tounicode_dict API can't auto-select — programmatic callers have to re-implement the scoring. (We do exactly that in the in-browser Easy Tibetan Copy worker.)

API

A new strategy= argument, backed by a resolve_strategy helper in tounicode_core:

patch_pdf(path)                                  # unchanged: default gid tier
patch_pdf(path, strategy="gid-pua-free")         # pin one bundled tree, no scoring
patch_pdf(path, strategy="auto")                 # score every bundled tree, keep best
patch_pdf(path, strategy=["gid", "gid-pua-free"])# score over an explicit subset

The subset form is the one that matters for us: a browser consumer can't afford to auto-try the 139 MB gshape tree, so "auto over these strategies" is what we need.

Design choices (open to your preference)

• Default unchanged. strategy=None keeps the historical behaviour (default gid tier) and adds no new key to stats, so existing callers are unaffected. I did not flip the API default to "auto" the way the CLI does — that felt like the maintainer's call given the docstring's "the public API is tier 1 (gid)" stance. Easy to change if you'd prefer auto-by-default.
• When a strategy is given, the winning name is reported as stats["strategy"].
• font_lookup_dir and strategy are mutually exclusive (raises ValueError).
• resolve_strategy reuses the existing _select_auto_strategy / _default_strategy_specs; single-name strategies skip scoring; unknown names and uninstalled trees raise clear errors.

Tests

tests/test_api_strategy.py covers: default-unchanged (no strategy key), pinned named tree, subset auto-select picking gid-pua-free on the #16 page, "auto", build_tounicode_dict(strategy=...), and the two error cases. Full suite: 81 passed, 10 skipped locally.

Stacks cleanly on top of #17 (independent file).