Support read_block with no delimiter and length=None

[adenzhou1350]

fsspec.utils.read_block documents and annotates length=None as reading through EOF. That works with a delimiter, but without one the function reaches assert length is not None:

from io import BytesIO
from fsspec.utils import read_block

read_block(BytesIO(b"abc"), 1, None)  # currently raises AssertionError; expected b"bc"

Use an unbounded read() for None, retaining read(length) for finite lengths. Add real-file and BytesIO regressions for start/interior/EOF/past-EOF offsets, with zero and finite-length controls, and clarify the docstring type. Existing delimiter handling is unchanged.

Validation against master 94afb3f6, Python 3.12:

• Unmodified Windows and Linux: 8 failures / 16 passing controls.
• Fixed Windows utility module: 157 passed.
• Fixed Linux utility, filesystem-spec, local and memory modules: 773 passed, 28 skipped, 1 existing xfail, 1 existing multithreaded-fork warning.
• Repository-pinned Ruff 0.14.3 check/format and diff check pass. Linux tested source bytes match the committed Git blobs.

Distinct from #2256 (delimiter + split_before) and #2261 (filesystem-level negative lengths). AI-assisted and independently reproduced. These are scoped native tests, not cloud-service or full-CI qualification.

[martindurant]

Non-test code should not contain asserts anyway (they can be stripped in bytecode) ...

[martindurant]

Isn't this just f.read(length), since it accepts None and does the right thing?

[adenzhou1350]

Yes, simplified to f.read(length) in 75a5878. Also resolved the changelog conflict with current master. Rechecked the final head: Windows utility/core tests 262 passed; Linux utility/spec/local/memory/core tests 932 passed (existing skips/xfail retained). Thanks!