When miniaudio.stream_any's decoder init issues a read larger than
(BUFFER_SIZE - position) while position is still inside the seekable
headroom, PatchedIceCastClient.read deadlocks against its downloader
thread:

- read() polls until len(buffer) >= num_bytes, where len(buffer) is
  size (post-position bytes).
- The downloader's _buffer.fits(BLOCK_SIZE) guard checks the *raw*
  buffer length, not size, so once raw is at BUFFER_SIZE it refuses
  to add even though size has room to grow with consumption.
- Reader waits for the buffer to grow; downloader cannot grow it
  until the reader consumes. Polling loop hits DEFAULT_TIMEOUT (10s)
  and miniaudio surfaces the resulting OperationTimeoutError as
  DecodeError('failed to init decoder', -1).

Small streams escape because the HTTP body completes within the 10s
window: _stop_stream flips True and the loop's other exit fires.

Detect 'raw buffer cannot grow' in the polling loop and let
self._buffer.get() return a short read. miniaudio's drmp3 handles
short reads by re-calling; once subsequent reads advance position
past HEADROOM_SIZE the buffer releases its headroom and becomes a
normal sliding window, after which the downloader can refill.

Tests use pytest-httpserver to serve the existing static_3sec.ogg
fixture (small, exercises the fast path) and a new audio_long.mp3
(~80 KiB silence, just over BUFFER_SIZE) to trigger the deadlock.
On master the long test fails after ~10s with DecodeError(-1); with
this fix both tests pass in under a second.

Also bump the miniaudio dev pin from 1.61 to 1.71 in
requirements/requirements.txt. The deadlock fires on 1.71's decoder
init call sequence (the new seek(0, END) plus extra round-trip
leaves position at 2048 when the read(65536) fires); 1.61's init
ends with seek(0, START) which resets position before the same
read, hiding the bug. Real users `pip install pyatv` and resolve
miniaudio>=1.45 to the newest available (1.71 today, and only 1.71
has Python 3.14 wheels), but the old dev pin meant pyatv's own CI
didn't exercise the bug pattern users hit. Bumping aligns CI with
what users actually run so the new regression test serves as a real
guard.