PERF: read_csv fast paths for nullable and pyarrow-backed EA dtypes

[jbrockmendel]

Requesting an ExtensionArray dtype explicitly (e.g. dtype="Int64", dtype="int64[pyarrow]", dtype="boolean") is currently 5–9x slower than the equivalent dtype_backend inference, because the c engine materializes an intermediate ndarray[object] of Python strings and then re-parses it (via to_numeric / _from_sequence_of_strings). This routes those dtypes through the nogil C converters instead.

• nullable Integer/Floating (e.g. Int64, Float64): parse with the same C converters dtype_backend="numpy_nullable" uses, building the NA mask from the actual NA positions rather than the integer sentinel, so legitimate int64-min / uint64-max values are preserved (the sentinel-based inference path masks them).
• pyarrow-backed numeric (e.g. int64[pyarrow]): the same C-converter parse, then wrapped as pyarrow — this keeps explicit int64[pyarrow] parsing identical to dtype_backend="pyarrow" rather than adopting pyarrow's more lenient string casting (which would e.g. accept "0x1F"). Temporal / boolean / string ArrowDtype build a pyarrow large_string array straight from the buffers and hand it to _from_sequence_of_strings (same conversion engine as before).
• nullable boolean (boolean): a nogil pass recognizing the true/false sets plus the "1"/"1.0"/"0"/"0.0" spellings, testing the true set before the false set to match map_string's ordering (so it stays correct even with overlapping user true_values/false_values).

Every fast path falls back to the generic string path on anything it cannot confidently handle (non-numeric data, ints wider than uint64, invalid UTF-8, an unrecognized boolean token, a dtype the requested type can't hold), so behavior — including raised errors — is unchanged. Also makes ArrowExtensionArray._from_sequence_of_strings compute its isna mask lazily so a pyarrow-array input isn't scanned unnecessarily.

Timings (1M rows × 10 cols, single-threaded), before → after:

dtype	before	after
Int64 / Float64	3419 / 3092 ms	399 / 614 ms
int64[pyarrow] / double[pyarrow]	3515 / 3268 ms	408 / 635 ms
bool[pyarrow]	811 ms	498 ms
boolean	1556 ms	312 ms

Validated with per-family differential tests (random + adversarial inputs, custom na_values, the parallel-read path) against the legacy path; whole result and raised errors match.