Treat FSMap keys and roots as literal paths

[kokotatan]

FSMap keys containing glob characters are currently interpreted as patterns when reading or deleting them. For example, with both chunk[0] and chunk0 present, reading chunk[0] returns the wrong file (and can return a dictionary), and deleting it can delete chunk0. The same issue affects a root such as store[0], including clear().

Use literal cat_file/rm_file operations for these paths, preserving existing bulk cat/rm calls for ordinary paths and rm-only backend compatibility for ordinary deletions. Literal bulk reads retain raise/omit/return error handling and duplicate-key behavior. Add regressions across memory, local, async wrapper, and simplecache, plus deletion/pop controls for an rm-only backend, and a changelog entry.

For roots containing glob characters, clear() removes all mapping keys while leaving any empty subdirectories. Bulk operations on paths containing glob characters run per file; ordinary bulk operations retain their existing implementation.

Validation (Windows, Python 3.13.11):

• Exact original mapping implementation: 32 new regression cases fail, 8 literal-path controls pass; final mapping suite: 77 passed, 95% branch-inclusive module coverage.
• Full suite before the final rm-only compatibility controls: 1,979 passed, 293 skipped, 11 xfailed, 2 deselected. The two deselections use | in local filenames and independently fail with WinError123 on unchanged tests.
• Final full suite: 1,977 passed, 293 skipped, 11 xfailed, the same 2 deselected; 4 unchanged GitHub integration tests fail with HTTP403 API rate-limit responses from mwaskom/seaborn-data (these passed in the earlier run). Optional dependency/cloud tests account for the skips.
• All prescribed pre-commit hooks passed; strict Sphinx HTML build passed.

AI assistance disclosure: implementation and tests prepared using OpenAI GPT-6; independent read-only review by Anthropic Claude Opus, with its compatibility finding reproduced and fixed. No human review is implied by this disclosure.

[martindurant]

From the very purpose of this PR, I think it is reasonable to assume that none of the inputs to a mapping have magic characters. Have you found usage where this is untrue?