Enable Parakeet V3 download and full Intel NPU execution

[derekszen]

Closes #12

Summary

Enable Parakeet V3 download and full Intel NPU execution by fixing two blockers:

• Point PARAKEET_V3 at the published FluidInference/parakeet-tdt-0.6b-v3-ov export instead of the missing 1.1b-v3-ov repository.
• Normalize negative opset5 LogSoftmax axes before compiling OpenVINO IR models on Intel NPU.

The joint graph uses axis=-1 on a static rank-4 tensor. This is valid OpenVINO and semantically identical to axis=3, but Intel's NPU compiler fails in AlignDimensionsForDPU, causing Eddy to run the joint component through its CPU fallback.

Implementation

• Correct the Parakeet V3 Hugging Face repository ID.
• Correct the README model-card link to the same published repository.
• Read non-blob IR models into memory when the requested device is exactly NPU.
• Visit opset5 LogSoftmax nodes with a negative axis and static input rank.
• Convert valid negative axes to their canonical non-negative value.
• Revalidate the graph only when a node changed.
• Leave CPU/GPU/AUTO compilation and compiled-blob imports unchanged.
• Leave dynamic-rank or out-of-range axes untouched so the existing compile failure and CPU fallback remain authoritative.

This follows the existing in-memory NPU graph-rewrite precedent in the Nemotron backend while keeping the workaround narrowly scoped to the affected operator and device.

Reproduction before the fix

Environment:

• Intel Panther Lake NPU5010
• Intel NPU driver 1.32.1
• OpenVINO 2026.2.1
• FluidInference/parakeet-tdt-0.6b-v3-ov

AlignDimensionsForDPU Pass failed : Got negative index -1 for Dim

The failing node is opset5 LogSoftmax logits, input shape [1, 1, 1, 8198], axis=-1.

Validation

• Built all Eddy targets successfully on Linux with OpenVINO 2026.2.1.
• Added openvino_utils_test with five hardware-independent cases: -1 -> 3 with CPU output equivalence, -4 -> 0, positive-axis no-op, dynamic-rank no-op, and non-LogSoftmax no-op.
• CTest: 1/1 passed.
• Confirmed the Parakeet V3 joint graph compiles and runs on NPU on an Intel Panther Lake NPU5010 with driver 1.32.1.
• Full 14.24-second sample: exact 83-token match against CPU and the previous hybrid path.
• Five speech slices: exact CPU/NPU transcript and token-count matches.
• Ten-second silence: empty on both CPU and NPU.
• Ten randomized joint inputs: identical winning token and duration heads on CPU and NPU.
• Median model inference: 196 ms with CPU joint fallback, 116 ms all-NPU, about 41% lower latency.

Scope

This does not modify model files on disk or broaden NPU fallback behavior. The graph workaround only canonicalizes a semantically equivalent axis representation before NPU compilation.