Remove the internal TOMLChar wrapper

[tfoutrein]

Stacked on #489, #490 and #491 — the capstone of that series; best reviewed/merged after them.
This also supersedes #488 (interning TOMLChar): per @dimbleby's suggestion on that PR, removing the wrapper entirely is the better end-state, so I'd close #488 in favour of this.

What

After the bulk run-scans (#490/#491), the parser only constructs a TOMLChar (a str subclass) at run boundaries and uses a handful of its is_*() helpers. This removes the class entirely:

• Source yields plain str characters; inc() / advance_* read self[i] directly.
• End-of-input is detected positionally (_idx >= len / Source.end()) instead of an identity sentinel.
• The remaining character-class checks use module-level frozensets.

A real NUL byte is still rejected as an invalid control char and is never mistaken for end-of-input, since EOF is now positional rather than a value/identity comparison.

Benchmarks

Median, interleaved A/B vs master (includes #489–#491):

document	speedup
large flat, single-line strings (~90 KB)	5.8×
poetry.lock-like (~64 KB)	2.4×
pyproject.toml	1.9×
typical mixed (~4 KB)	1.6×

The removal itself adds ~1.1–1.18× over #491. No regression on any shape.

Tests

Full suite passes (972, incl. the toml-test conformance submodule). On top of that, an 11.5k-input adversarial differential — EOF/truncation at every prefix length, real-NUL placement in every position, empty/whitespace/BOM, and structural fuzz — is byte-identical in output and exception type to master. No public API change (TOMLChar was not exported).

[tfoutrein]

Rebased onto master now that #491 has merged — single commit on top of current master, no conflicts (composes cleanly with the recent escape/bare-key fixes #493/#497/#501). CI is green. This removes TOMLChar entirely, so it supersedes #488 (interning) — I'll close #488 once this lands. Ready for squash-merge. Thanks!

[tfoutrein]

Done — dropped the duplicate raw-string constants and kept only frozensets, renamed without the _SET suffix (your preferred option). The two consume(...) calls that need a plain string now pass a literal. No behaviour change; the full test suite + conformance still pass.

Note on CI: all unit-test jobs (every OS × Python) + pre-commit + the poetry integration are green. The two poetry-core integration jobs failed at poetry install with a vendoring/lockfile resolution error ("This is likely not a Poetry issue…") — unrelated to this PR (it passed on master and on the Windows poetry-core job). I don't have permission to re-run them on this repo, but a re-run should clear it.