Enable compressonator-bc7e on macOS

[JorgeVV]

Makes the compressonator-bc7e backend available on macOS. It was hard-gated off because no darwin build of the fork existed — §8 of the fork's README lists macOS as out of scope, and its v1.0.0 release ships Linux and Windows binaries only.

Companion PR: noisethanks/compressonator#1 — the build script and the source fixes that make a macOS binary possible. Merge that one first; this PR embeds its output.

What ships here

internal/tools/bin/compressonator-bc7e-macos — universal (x86_64 + arm64), ad-hoc signed like texconv-macos and 7zz-macos, 6.1 MB. Everything else is un-gating: embed_darwin.go embeds it, config.normalizeBackend stops coercing darwin to texconv, and the Settings row becomes visible. Availability still keys off EmbeddedTools.CompressonatorPath == "" rather than runtime.GOOS.

Native Apple Silicon releases

Second commit. The release built darwin/amd64 only, so every Apple Silicon user who downloaded a release ran under Rosetta 2 — and so did every embedded tool, because a universal child process inherits the parent's slice. Verified directly:

parent	universal child runs as
arm64	arm64
x86-64	x86-64

So texconv-macos, 7zz-macos and the new compressonator-bc7e-macos were all translated despite shipping native arm64 code. macOS now ships two archives:

• atak-<version>-macos-x64.tar.gz
• atak-<version>-macos-arm64.tar.gz

Two archives rather than one universal binary: the tools are embedded per GOOS and each already carries both slices, so a universal atak would hold two complete copies at ~45 MB against the 25 MB target. Each single-arch build is 22 MB (9.2 MB compressed). Verified with a full goreleaser release --snapshot — four binaries, four archives, correct names, Mach-O 64-bit executable arm64 inside the arm64 tarball.

Tests now run on macOS

dispatch_srgb_test.go and e2e_srgb_summary_test.go hardcoded /home/abhi/... paths and skipped on every other machine. They now resolve the binary from internal/tools/bin for the current platform, so TestDispatchCompressonatorSrgbFallback exercises the compressonator → texconv fallback on macOS and passes.

Verification (Apple Silicon, M4 Max)

Check	Result
Format sweep, both slices	38/40 configs; the 2 failures are sources already in the destination format, identical on both slices
Real GAMMA textures, uncompressed sources	38/39 (ATAK skips compressed sources at scan time, so this is the real input)
End to end via tools.Extract + RunPool	12/12, byte-identical across two full runs
Quality vs stock BC7	within ±0.7 dB either way
Speed vs stock BC7 at -Quality 1.0	0.54 s vs 24.70 s, ~46×

Caveats worth reviewing

1. Cross-platform output is no longer bit-identical. The arm64 slice differs in the low bits from the x86-64 builds on 18 of 38 configs, at equal PSNR. Output stays reproducible within a slice. SPEC.md says so explicitly; the previous "bit-identical output" claim was reworded rather than quietly dropped.
2. Release size: macOS goes ~17 MB → ~22 MB per architecture, now the tightest of the three against the 25 MB ceiling.
3. .goreleaser.yaml still uses archives.format and archives.builds, both deprecated. That is pre-existing — I matched the file's existing style rather than mixing dialects — but the workflow pins version: latest, so a future goreleaser release could break it. Worth a separate cleanup.

Not changed, on purpose

One real texture in the sample fails with unsupported input image file format, which does not match the narrow Could not load source file fallback pattern, so no texconv retry happens. That behaviour is pre-existing and identical on all platforms, so widening the matcher did not belong in this PR.