Reduce CPU usage on ARM (AArch64 & ARMv7)

[rikkus]

PR against Core
PR against Plugin

These optimisations provide:

AArch64

>= 2x speedup vs a2_fast

Tested and benchmarked on

• Apple Silicon M2 (Macbook Air)
• Cortex-A76 (Raspberry Pi 500)

ARMv7

>= 1.3x speedup vs a2_fast

Tested and benchmarked on

• ARM Cortex-A17 (RK3288, ASUS Tinker Board, same SoC as HeadRush Core & Prime)

Bit-identical to a2_fast as tested - 64 and 32 frame blocks

A separate repo used for working on these optimisations runs some build & conformance checks.

---

Tested against Plugin PR #679 (f1a8c79) with Core at 552c5ab (head of sdatkinson/NeuralAmpModelerCore#313), on an Apple M2, macOS 27 beta, Xcode 27 beta. Used a local MacOS 12.0 override to make Xcode build - override not added to any PR.

• [✅] The standalone plugin builds. (Universal arm64 + x86_64.)
• [✅] The plugin runs. (Standalone app, and the AU in Logic Pro.)
  • [✅] The plugin makes sound.
  • [✅] You can load a new-style (file) model. (A2 captures, both submodels via the Slim knob.)
  • [✅] You can load an old-style (directory) model. I chose a directory in the app - is that what this means? It wouldn't let me choose a directory in the plugin.
  • [✅] You can load a supported IR.
  • [✅] The EQ section works.
• [✅] The VST3 plugin builds and can be loaded by the VST3 SDK. N.B. I loaded it in Logic Pro, not with the validator.
  • [✅] The plugin passes all unit tests implemented by the VST3PluginTestHost's unit testing tool. (Ran the same suite with the SDK's command-line validator -e: 537 passed, 0 failed.)
• [✅] The AU plugin builds. (Also passes auval -v aufx 1YEo SDAa in full, including its render tests.)

Not tested: Windows (no machine available) and iOS (checked at the project-file level only).

[rikkus]

@sdatkinson I think I'm finally at a good point with this optimisation work, with ARMv7 now added, full (as much as I can reach) test coverage and proper benchmark publishing. I hope it's useful and doesn't have fatal flaw[s]!