[L0FKA]

Apply an audio taper to endpoint volume

Rocksmith adjusts guitar input gain through IAudioEndpointVolume::SetMasterVolumeLevelScalar: 0.17 at start, a
0.40..0.50 search during calibration, then 1.0. RS_ASIO uses that scalar directly as a linear amplitude
multiplier. A Windows endpoint scalar follows an audio taper, so the same setting gives very different
attenuation: the linear multiplier is +11.4 dB too hot at 0.17 and +4.5..5.9 dB too hot at 0.40..0.50.

What changes

• Endpoint taper: the endpoint scalar maps to amplitude as scalar^1.74, fitted to measurements of a real
  Windows software endpoint (MOTU M Series), within 0.3 dB of every measured point. At 0.17 the attenuation
  becomes -26.78 dB instead of -15.39 dB.

• Unity stays unity: the taper does not change the gain at scalar 1.0, which stays 0 dB. It does not fix
  distortion while the endpoint is at full scale. SoftwareMasterVolumePercent keeps its linear meaning.

• Integer software volume: Int16, packed Int24 and Int32 retain whole-percent integer scaling in the callback.
  This avoids converting each sample to floating point and back on a performance-sensitive path. The 1% step is
  an intentional trade-off; integer tenths-of-a-percent scaling can be considered if testing shows it is needed.

• Endpoint API: implements:

  • master volume in dB, get and set;
  • the range (-96..0 dB, 1.5 dB steps), with no digital gain above unity, since the player sets the level with
    the interface's analog gain;
  • channel count;
  • volume steps;
  • QueryHardwareSupport (0, software only).

  Invalid and NaN settings are rejected. With software endpoint volume disabled, the new queries report
  E_NOTIMPL, so no range is advertised that the setters would refuse. Mute, per-channel volume and
  notifications stay unimplemented.

• Audio path: the taper is computed in the setter and cached. The audio callback reads the cached amplitude
  without locking. Setters are serialized, so the reported level and the applied gain can't drift apart under
  concurrent calls. Applies to capture and render.

Testing

• Offline: standalone taper checks covered the measured curve, round trips, the floor, dB quantization and
  steps. A deliberately wrong exponent fails them. Release x86 builds.

• Live capture readout (temporary instrumentation in a validation-only build, not part of this PR):
  September 2022 build, MOTU M Series Int32LSB, endpoint scalar 0.17.

  Build	Before volume	After	Gain
  master	-12.00 dBFS	-27.39	-15.39 dB (linear)
  this PR, whole-percent integer DSP	-12.00	-39.96	-27.96 dB

Not established: improved calibration or scoring with a real guitar, long sessions, or equivalence to every
device's taper. This is one software-endpoint curve, measured on one interface. It does not change the
interface's analog gain.

[mdias]

The reason integer volume is applied in whole percents instead of precisely is performance. Being off by 1% is not going to get noticed, but if we're at the limit of what the system can push, converting the samples from integer to floating point, multiplying and then converting back to integer might be the thing that generates an xrun.

The whole-percent math is preferable. And if 1% is not enough precision we could multiply and divide by 1000 to get 0.1% accuracy.