Fix float-to-integer PCM conversion in to_wav (int8 crash, norm=False silence, truncation bias)

[gaoflow]

MicrophoneArray.to_wav converts a float signal to the requested bitdepth with a bare
np.array(signal, dtype=bitdepth), i.e. no float-to-PCM conversion at all. That breaks the
write side of the same dtype handling #440 (read) and #443 (normalize) already fixed:

• norm=False + any integer bitdepth silently writes an all-zero file. Simulation signals are
  floats in ~[-1, 1]; the bare cast truncates every sample to 0. No error, just silence on disk.
• bitdepth=np.int8 always raises ValueError: Unsupported data type 'int8'. WAV 8-bit PCM is
  unsigned; the documented int8 option can't work with a signed cast. This is the write side of
  the unsigned-8-bit format Fix to_float32 producing nan/inf for unsigned-integer (8-bit WAV) input #440 taught to_float32 to read.
• Truncation instead of rounding on every float->int cast gives a sign-dependent ~1 LSB bias
  (mean -0.5 LSB for x>0, +0.5 for x<0). This hits the documented happy path too
  (examples/room_from_rt60.py uses norm=True, bitdepth=np.int16), and the exported to_16b.

#440 made an 8-bit WAV readable; this makes one writable. The conversion is now the exact inverse of
to_float32: scale by full scale 2**(bits-1), round, clip so the cast can't overflow, and write
int8 as unsigned offset-binary. So to_wav -> wavfile.read -> to_float32 round-trips.

Round-trip residual (write -> read -> to_float32, 1 kHz tone), before vs after:

bitdepth   norm     before                 after
int8       False    ValueError             r=1.000 (uint8 on disk)
int8       True     ValueError             r=1.000 (uint8 on disk)
int16      False    all-zero file          resid 2.0e-05
int16      True     resid 2.85e-05         resid 1.6e-05   (rounding)
int32/64   False    all-zero file          resid ~3e-08
to_16b     n/a      err [-1, +1] LSB       err [-0.5, +0.5] LSB, no bias

One judgment call: norm=False + a float signal + integer bitdepth goes from "all-zero garbage" to
"input assumed full-scale [-1, 1], clipped" (the soundfile/librosa convention). Nothing sane consumes
the current all-zero output, but if you'd rather raise for out-of-range input than clip, happy to
switch that hunk.

Tests in tests/test_to_wav.py cover every bitdepth x norm round-trip (including the previously
crashing int8 and all-zero norm=False rows), the uint8-on-disk / silence->128 behavior,
full-scale saturation without wraparound (int64 is the sharp case: 2**63-1 isn't float64
representable), rounding, multichannel, integer-input passthrough, and the to_16b bias. Ran
tests/test_to_wav.py plus test_to_float32, test_normalize, test_microphone_array,
test_metrics, test_create_noisy_signal locally (59 passing); did not run the heavy
simulation/DOA suites.