[NPU] avoid HiFT linear interpolation device fault

[TKevin0v0]

[NPU] avoid HiFT linear interpolation device fault

Motivation

CosyVoice3 HiFT inference fails on Ascend NPU with a 507015 AI Vector
exception. Device slog identifies the first failing kernel as
UpsampleLinear1d_*_high_performance_2 with VEC instruction error: the ub address out of bounds.

The failing call is the SineGen2 phase downsampling used by 24 kHz
CosyVoice3:

input=(1, 9, 6720), float32, scale_factor=1/480, output=(1, 9, 14)

Making the input contiguous does not avoid the device fault.

Modifications

• Keep the existing native F.interpolate(..., mode="linear") path for CPU,
  CUDA, and other devices.
• On NPU, express the same align_corners=False source coordinates using
  index_select and interpolation weights.
• Keep the entire fallback computation on the NPU; there is no CPU compute
  fallback.
• Add coordinate-equivalence tests for the real 6720 -> 14 downsampling and
  the reverse 14 -> 6720 scale.

Correctness and runtime validation

• 6720 -> 14: max absolute difference vs. CPU reference = 0.
• 14 -> 6720: max absolute difference = 1.19e-7.
• Standalone HiFT with a real Flow mel tensor succeeds and returns
  (1, 6720) float32.
• SGLang-Omni /v1/audio/speech returns HTTP 200 and a valid PCM16 mono
  24 kHz WAV.
• Focused unit tests: 2 passed.
• Repository flake8 configuration and git diff --check: passed.

Scope

This is a model-side workaround for an Ascend built-in operator fault. It does
not claim to patch the internal CANN kernel implementation.

Related: sgl-project/sglang-omni#1652