[PERFORMANCE] Optimize long linear convolution scheduling

[sdatkinson]

Summary

• replace the uniform 1,024-sample tail with exact non-uniform power-of-two convolution tiers
• distribute spectral multiply work over each tier deadline and phase-stagger transforms
• use real half spectra and a benchmark-tuned dispatch table from 1K taps through one-minute responses
• add callback-distribution benchmarking, full-length impulse verification, and dispatch/correctness tests

This is stacked on #323 because the plugin update currently consumes that Sequential branch.

Performance

On Apple M1, Release -O3, 48 kHz, 32-sample callbacks:

• 1,208,121 taps: p99 107 us, max 525 us, 0 overruns (666.67 us deadline)
• 2,880,000 taps: p99 112 us, max 325 us, 0 overruns
• previous atmospheric-model implementation: approximately 2.7-3.1 ms p99

The dispatch choices and candidate tradeoffs are recorded in tools/BENCHMARK_LINEAR.md.

Validation

• full Core Release build
• complete Core test suite
• AddressSanitizer + UndefinedBehaviorSanitizer test suite
• allocation-tracking tests
• full 2,880,000-tap impulse render: max absolute error 6.4e-10
• full 1,208,121-tap impulse render with 63-sample callbacks: max absolute error 6.4e-10
• universal arm64/x86_64 Release VST3 build and local install