Optimize long linear convolution scheduling

[sdatkinson]

Summary

• update NeuralAmpModelerCore to [PERFORMANCE] Optimize long linear convolution scheduling NeuralAmpModelerCore#324
• replace the periodic long-filter CPU burst with benchmark-tuned non-uniform, deadline-distributed convolution
• preserve exact zero-latency output and real-time-safe steady-state processing

This is stacked on #670.

Measured result

For the 1,208,121-tap atmospheric model at 48 kHz with 32-sample callbacks, Release -O3 callback p99 fell from approximately 3 ms to 107 us; the measured run had no 666.67 us deadline overruns.

The complete one-minute 2,880,000-tap impulse response matched with 6.4e-10 maximum absolute error.

Validation

• Core Release and sanitizer test suites
• allocation-tracking tests
• benchmark sweep from 1,024 through 2,880,000 taps
• universal arm64/x86_64 Release VST3 built and installed locally