Add Sqeakd direct-ANE research harness

[ebowwa]

What changed

• integrates upstream PR ANE private API research: chaining, E5 runtime, custom MIL compilation maderix/ANE#40's custom MIL/E5 research suite
• preserves current main's compiler and security flags while resolving its Makefile conflict
• adds the missing Core ML fixture generator required by the suite
• adds a Sqeakd MLProgram graph inventory command
• documents the direct-ANE validation results and correctness-first porting order

Why

Core ML reports the fixed Sqeakd FP16 graph as ANE-compatible but schedules it
on GPU or CPU. This branch establishes a macOS-only private-API research path
for translating and benchmarking Sqeakd kernels directly on ANE.

Validation

• make test_mil_custom
• ./test_mil_custom on M1/macOS 26.5.2
• attention, linear, transformer block, and backward matmul CPU-reference checks pass
• make sqeakd-inspect MODEL=/path/SqeakdCTC_fixed105_fp16.mlpackage
• graph inventory reports 1,877 operations and the expected fixed tensor contract

Scope

This is research code using private APIs. It is intentionally isolated from
the iOS, TestFlight, App Store, and production Sqeakd runtimes.