Bound the buffered read to the range the server declared

[Gares95]

Description

PartialBuffer represents a declared byte range of a remote file, but on the
non-stream path it reads the whole response body:

self.buffer=bufferifstreamelseio.BytesIO(buffer.read())

buffer.read() with no argument consumes whatever the server chooses to send,
and it happens before size is consulted at all. So how much gets buffered is
decided by the server rather than by the range requested. Asking for a 100 byte
range from a server that streams 20 MB buffers all 20 MB:

server sends	range requested	bytes buffered
100	100	100
1,000,000	100	1,000,000
20,000,000	100	20,000,000

That defeats the point of using range requests, and for anything that runs
remotezip against a URL it does not control it is a way to exhaust memory
remotely.

There is a second route to the same place. fetch turns the server's
Content-Range straight into the size without checking it:

range_min, range_max=self.parse_range_header(range_header)
returnPartialBuffer(res, range_min, range_max-range_min+1, stream)

A header whose end precedes its start gives a negative size:

>>>RemoteFetcher.parse_range_header('bytes 100-50/1000')
(100, 50) # size becomes 50 - 100 + 1 = -49

PartialBuffer.read(0) then computes a negative length, and file.read(negative)
means read everything for a Python file object. End to end, a client that asked
for 100 bytes received the entire 5000 byte body.

Separately, a header that cannot be parsed leaks a bare ValueError out of the
library: 'bytes abc-def/1000', 'bytes -500-100/1000', 'bytes /1000' and an
empty value all do this today. So does 'bytes */1000', which is the
RFC 7233 unsatisfied-range form, so this is not only about hostile input: a
server can send that legitimately.

The change

• Read at most size bytes in PartialBuffer, in a loop that tolerates short
  reads. A single read(size) call is not sufficient: a socket-backed
  response can return fewer bytes than requested while more are still coming, so
  reading once would silently truncate. Worth stating because the existing tests
  all use BytesIO, which never short-reads, so that failure mode is invisible
  to them. The data is written straight into the result buffer, so no
  intermediate copy of the whole range is held.
• Raise RemoteZipError for a Content-Range that ends before it starts or
  carries no end, and for one that cannot be parsed at all.

The suffix form that parse_range_header returns for bytes -123 is left
alone, since that is a request form rather than a response header, and its
existing test still passes. bytes 0-99/*, an unknown total length, remains
accepted and has a test to keep it that way.

What this does not fix

Worth being explicit: this bounds buffering by the size the server declares,
not by the size the client requested. A server that declares a very large
range and then streams it will still be buffered in full:

server declares	server sends	buffered
bytes 0-99/1000	5,000,000	100
bytes 0-4999999/5000000	5,000,000	5,000,000

Closing that would mean rejecting or clamping a response range that does not
match the requested one. I have not done it here because the right policy is a
judgement call: some servers legitimately return a different range than asked
for, and RemoteIO.seek derives _file_size from the declared size, so
clamping naively would desynchronise it. Happy to follow up if you have a
preference.

Tests

Four added. The first two fail against the current behaviour:

• a server sending more than it declared does not enlarge the buffer;
• an invalid or malformed Content-Range is rejected.

The other two guard the new code rather than the old bug, and pass before the
change as well, since reading to EOF never truncated:

• short reads from the stream are handled without truncation;
• a server sending less than it declared does not hang or raise.

Validation

python test_remotezip.py goes from 19 tests to 23, all passing, with no
existing test modified.

[gtsystem]

Hi, thanks for contributing.
I added one comment. If you can also bump the minor version would be great.

[gtsystem]

_read_up_to() now stops after size bytes, so the original HTTP response may remain partially consumed. Since buffer is discarded here, its connection is never explicitly closed or released. This can exhaust the HTTP connection pool when a server sends more bytes than declared. Please close the source in a finally block after copying, including calling release_conn() when available.

Example:

ifstream:
self.buffer=bufferelse:
try:
self.buffer=self._read_up_to(buffer, size)
finally:
buffer.close()
ifhasattr(buffer, 'release_conn'):
buffer.release_conn() # release urllib3 connection associated with this buffer