AOT GGUF converter for I2_S — eager-exec half of #1207

[michalharakal]

Sub-issue of #1198. Eager-exec half of #1207 (the IREE/.irpa leg is a separate follow-up, not
addressed here).

Why this instead of #1204's sidecar

The on-device sidecar cache (#1204) pays its "convert once" cost by doubling the first cold
start — decode + re-encode + write-to-disk — on the exact constrained device this effort exists
to protect. An AOT converter that runs once, offline, ahead of the device ever seeing the file,
avoids that entirely: a build that controls its own model pipeline gets a genuinely
zero-runtime-cost load, no sidecar, no first-load penalty. Full reasoning on #1198/#1207.

What this PR does

• GgufTensorEntry gains rawBytes: ByteArray? as an alternative to tensor: a converter
  that already has the exact on-disk bytes for an entry (a passthrough of a quantized blob this
  writer doesn't need to interpret, or a buffer whose true size exceeds its type's formal
  GGML_QUANT_SIZES block math — an I2_S BITNET_B1_58 buffer with its trailing FP32 scale) can
  hand them over directly. GGUFWriter's expectedTensorSize/materializeTensor use
  rawBytes.size and the bytes verbatim when present, bypassing TensorFlatten (element-indexed,
  can't represent an opaque quantized blob) and GGML_QUANT_SIZES entirely. Every existing
  tensor-mode caller/test is unaffected — tensor stays the default path.
• I2sScale.kt: extracted StreamingGgufParametersLoader's I2_S scale-resolution logic
  (resolveI2sScale, i2sCompanionScaleTensor, i2sTrailerScale, i2sTrailerScaleIsMappable)
  into shared functions, so the converter gets the identical answer the streaming loader would,
  rather than a second, potentially-drifting copy.
• I2sAotConverter.convert(): reads an arbitrary GGUF via StreamingGGUFReader, repacks
  I2_S tensors into SEQUENTIAL+trailer order (I2sRepack — the same repack the loader already
  does, just ahead of time), drops the now-redundant companion scale tensor (keeping one would
  defeat Skip the I2_S repack copy for SEQUENTIAL-layout GGUFs, enable true mmap #1203's mmap fast path, which requires no companion to exist), and passes every other
  tensor and all KV metadata through unchanged. Produces a GgufWriteRequest for GGUFWriter.

Test plan

• skainet-io-gguf full JVM test suite passes (including the pre-existing
  GGUFWriterRoundtripTest — unaffected by the GgufTensorEntry change)
• New I2sAotConverterTest: converts a GROUP_128 (BitNet.cpp) source file, asserts the
  converted I2_S tensor decodes identically to the source and actually takes Skip the I2_S repack copy for SEQUENTIAL-layout GGUFs, enable true mmap #1203's mmap
  fast path afterwards (packedStorage is Storage.OffHeap under WeightResidency.MAPPED) —
  proving this is a real substitute for the on-device cache, not just the same cost moved
  earlier. A passthrough Q4_0 tensor is asserted unchanged by the round trip.