Mirror readinto from the pyarrow stream onto ArrowFile

[hampsterx]

ArrowFile mirrors a fixed method list from the pyarrow stream it wraps, and readinto is not on
it, though pyarrow.NativeFile implements it. That makes it the one file class here without one:
AbstractBufferedFile implements readinto (fsspec/spec.py), so anything written
against the usual fsspec handle contract breaks on an Arrow-backed filesystem and nowhere else.

LibArchiveFileSystem is the in-repo case, and it fails silently. custom_reader documents that
"the file object must support the standard readinto and 'seek' methods" and calls it from a
ctypes callback. An AttributeError there is swallowed by ctypes, libarchive sees a zero-length
read, and ls() returns an empty archive rather than raising:

>>> fs = ArrowFSWrapper(LocalFileSystem())
>>> LibArchiveFileSystem(fo=fs.open("a.tar", "rb")).ls("", detail=False)
Exception ignored on calling ctypes callback function ...
AttributeError: 'ArrowFile' object has no attribute 'readinto'
[]                      # on master; ['hello.txt'] with this patch

(The swallow itself is pre-existing and independent of this patch: any readinto failure in that
callback yields an empty archive rather than an error. This just removes the trigger.)

Buffered reads raise properly at least:

>>> io.BufferedReader(fs.open("a.txt", "rb")).read(3)
AttributeError: 'ArrowFile' object has no attribute 'readinto'

A whole-file read() is fine, since that path goes to read() on the raw object. A sized read,
peek() and read1() each fill the buffer through readinto.

Gzip is how it turned up. fsspec/compression.py registers isal's IGzipFile as the gzip codec
when isal imports and the stdlib GzipFile only when it does not, and isal decompresses through
readinto. So a .gz file read through any Arrow-backed filesystem fails in an environment that
merely carries the package, and reads fine everywhere else. A transitive xopen dependency brought
isal in on x86-64 and AArch64, and every .gz read started raising.

One entry on the mirror list, following #1154 (seekable) and #1944 (size).

Tests

test_readinto covers the handle directly on both the seekable and non-seekable paths;
test_readinto_supports_a_buffered_reader is parametrized over read(3), peek(3) and read1(3).
All eight cases fail on master with the AttributeError above and pass with the one-line change.

Neither the gzip nor the libarchive case is asserted: isal is not guaranteed on CI and libarchive is
optional, and a combined libarchive + pyarrow fixture does not exist here. Both were reproduced by
hand against master and this branch. One note for anyone reproducing the buffered cases: the
pure-Python _pyio.BufferedReader uses raw.read for sized reads and so does not show the bug,
only the C io.BufferedReader does.

A downstream workaround, for anyone hitting this before a release: panodata/omniload#314 mixes a
_open override into its Arrow-backed filesystem classes to expose the stream's own readinto.