Raw_New borrows a reference to its PyBytesObject argument, so Raw_Copy must
still dispose of it.
The following is a minimal repro for this issue:
import msgspec
DECODER = msgspec.msgpack.Decoder(type=msgspec.Raw)
PAYLOAD = b"\xda\xff\xff" + (65535 * b"\0") # array of 65535 zeros
for _ in range(1000000):
  raw = DECODER.decode(PAYLOAD)
  raw.copy()

Fixes msgspec#1108.
`Ext_New` doesn't steal the reference to `data` (it does its own
`Py_INCREF`, matching the borrow contract expected by the `Ext`
constructor). Both decode branches in `mpack_decode_ext` (typed `Ext`
decode, and `Any` decode without an `ext_hook`) passed a fresh reference
from `PyBytes_FromStringAndSize` and never released it, leaking one
bytes object (the ext payload) per decoded Ext. The `datetime` (code -1)
and `ext_hook` branches were not affected.
The fix releases the caller's reference after constructing the `Ext`,
same pattern as the `Raw.copy()` leak fix (msgspec#709). Verified empirically:
RSS delta over 200k decodes of a 1 KiB payload went from ~205 MiB to 0
on both affected branches; the new test fails before the fix
(`sys.getrefcount` 3 vs 2) and passes after.
Thanks @K-ANOY for the precise report, the ownership analysis there was
spot on.
Co-authored-by: Siyet <Siyet@users.noreply.github.com>