feat(compile-hlo): array-free external constants via BufferHandle.Floats, Long-safe sizing (#1247)

[michalharakal]

Part of #1247. Stacked on #1250 (needs the aliased FloatArray reaching the converter) — after #1250 merges, this retargets to develop.

Root cause found while implementing: the gemma3n tied embedding is 262144 × 2048 FP32 = Int.MAX_VALUE + 1 bytes — one byte over the JVM single-array ceiling. ByteArray(count * 4) overflowed to a negative size, and the resulting NegativeArraySizeException escaped into the converter's generic catch, producing the misleading Unsupported op 'weight' … Known names: […] comment the issue quotes. No widening of the byte round-trip could ever emit this tensor; the external path must be array-free.

• New sealed variant BufferHandle.Floats(FloatArray) (lang-core): FP32 logical values as a primitive array — up to 2 Gi elements / 8 GiB logical, read-only, typically aliasing the live module weight. DefaultBufferResolver reads it through a chunked little-endian byte view; describe labels it.
• External FP32 constants in ConstantOperationsConverter hand the aliased array to the packager as Floats with no byte serialization — zero copies from module weight to archive. Non-FP32 / under-filled values keep the padded byte path.
• IrpaWriter streams Floats little-endian in 64 MiB chunks — byte-identical output to the eager Owned serialization (parity-tested).
• elementCountFromShape folds in Long; oversized single-buffer serialization throws ConstantTooLargeException with the remediation — deliberately not IllegalArgumentException, so the inline-emission fallback can't catch it and try to format a multi-GiB tensor as text.

Testing: new BigConstantMaterializationTest (Long fold incl. the exact embedding count, overflow refusal, Floats identity-aliasing with zero serialization, tied-weight → one util.global) + IrpaFloatsParityTest. lang-core / compile-hlo / iree-params suites green; apiDumps updated.

Note for consumers: exhaustive whens over BufferHandle gain a branch — in-repo consumers are updated here; SKaiNET-transformers' Gemma3nExportHarness.writeSafetensors needs a Floats branch (chunked bf16 streaming) once this ships.