Fix percent-encoded base64 data URL reads

[bIackr0se]

Reading data:;base64,%5A%6D%39%76 returns six incorrect bytes instead of b"foo". Escaped padding and + or / characters can also cause padding errors.

Percent-decode the URL body before base64 decoding. This follows the decoding order in the WHATWG data URL processor.

with fsspec.open("data:;base64,%5A%6D%39%76", "rb") as f:
    print(f.read().hex())
# Before: e40e83dfdefa
# After:  666f6f (b"foo")

The tests cover escaped padding, alphabet characters, literal +, all 256 byte values, file size and byte ranges. Unescaped base64, empty bodies and decoded bytes containing a literal percent sequence have compatibility checks.

Verification on macOS arm64 with Python 3.14.7:

• Six regression cases fail against the upstream decoder and pass with the fix.
• Data, core and API tests: 105 passed, 6 skipped, 1 expected failure.
• All repository pre-commit hooks passed.
• A built wheel installed in a fresh environment passed 11 consumer cases, including binary reads, copies to local files, metadata, ranges and a UTF-8 text read.