fix: use rounding for float-to-integer conversions

[roderickvd]

This PR replaces truncating casts with proper rounding in float-to-integer sample conversions to eliminate systematic bias and nonlinear distortion.

Problem

The current implementation uses truncating casts (e.g. as i16), which creates two issues:

1. Nonlinear distortion: All signal values in the interval (-1.0, 1.0) map to zero, creating an output bin twice as large as any other integer value. This violates the uniform quantization assumption and introduces harmonic distortion.

2. Systematic bias towards zero: Small signals that should map to ±1 are instead lost to zero, introducing DC bias and reducing effective dynamic range by about 8 dB.

One publication that documents this is Dannenberg's "Danger in Floating-Point-to-Integer Conversion" letter to Computer Music Journal in 2002, which warns against truncation in audio applications.

Solution

Replace (s * scale) as {integer} with (s * scale).round() as {integer} for float-to-integer conversions.

Before (truncation):

s to_i16 { (s * 32_768.0) as i16 }  
// 0.1 * 32768 = 3276.8 → 3276 (truncated)
// 0.00002 * 32768 = 0.65536 → 0 (small signal lost)

After (rounding):

s to_i16 { (s * 32_768.0).round() as i16 }  
// 0.1 * 32768 = 3276.8 → 3277 (rounded)
// 0.00002 * 32768 = 0.65536 → 1 (small signal preserved)

The performance impact is minimal, because LLVM generates efficient code for round() intrinsics with dedicated instructions on most targets.

[roderickvd]

The failing build is unrelated and already present on master since cd8d893.

➡️ PR #192

[roderickvd]

Rebased and passing the CI now.