fix(backends): recognise counters the same way on every backend

[allen0099]

Closes #111.

Problem

Reproduced on all three backends before the change:

memory    increment(response '42') -> 43           increment(set counter_entry(5)) -> 6
redis     increment(response '42') -> CacheXError  increment(set counter_entry(5)) -> CacheXError
memcached increment(response '42') -> CacheXError  increment(set counter_entry(5)) -> CacheXError
decode_entry(b' 7 ') -> counter 7
decode_entry(b'1_0') -> counter 10

After the change, every backend gives CacheXError for the response and 6 for the counter, and both stored values decode to None.

Change

• counter_value() (types.py) now requires fingerprint == COUNTER_FINGERPRINT as well as an integer body. The memory backend and the base fallback therefore refuse a cached response like the network backends do.
• New parse_counter() (types.py) replaces int() in both counter_value and the codec. It accepts an optional -, ASCII digits and trailing spaces. It rejects leading whitespace, _, +, newlines and non-ASCII digits.
  • This differs from the issue's ^-?\d+$. Memcached pads a value that DECR shortens rather than resizing it: DECR on 10 leaves b"9 " (checked against a live server). A strict pattern would break every Memcached counter after a shrinking decrement.
• encode_entry() stores exactly what counter_entry(n) builds as a bare integer, which is the form INCR/INCRBY work on. This is the issue's optional part. Without it, the set()-then-increment() inconsistency would remain. An entry with a different fingerprint, a media type, or a non-canonical body such as 042 is still stored as a JSON document, and it round-trips unchanged.

These edge cases are the only behaviour change; nothing in the API changes.

Tests

• New tests/backends/test_counter_contract.py runs the same assertions on memory, the base fallback, Redis and Memcached:
  • a numeric response is refused and left intact;
  • set(counter_entry(5)) continues with increment();
  • a shrinking counter reads back as a counter.
    It also covers parse_counter and counter_value directly.
• test_codec.py:
  • the case " 7\n", which used to be accepted, is now rejected;
  • new tests cover the bare-integer encoding and the entries that must stay JSON documents.

Mutation checks, each with the rest of the fix in place:

Mutation	Failing tests
Drop the fingerprint check	3
Drop the bare-integer encoding	3 (including Redis and Memcached contract tests)
Go back to int()	12

The full suite passes against live Redis and Memcached (CACHEX_REQUIRE_LIVE_SERVERS=1): 826 passed, 100% coverage. mypy --strict is clean.

Like #156 and #157, this adds a ### Fixed section to ## [Unreleased], so the later merges need a small CHANGELOG rebase.