fix(voice): decode provider PCM16 as little-endian

[harshitethic]

Summary

Decode provider PCM16 bytes with an explicit little-endian dtype instead of the host-native np.int16 interpretation, then normalize to native int16 before returning audio to callers.

This fixes the big-endian-host corruption described in #4816 while preserving the existing np.int16 / np.float32 output shapes and odd-byte padding behavior.

Tests

Adds focused coverage using b"\x01\x02\x03\x04", which must decode to samples [513, 1027] for both integer and float output paths.

Closes #4816

AI assistance disclosure: AI assistance was used to inspect the issue, prepare the focused implementation and regression tests, and review the resulting diff.

[seratch]

Thanks for the contribution. #4817 already addresses #4816 with the same little-endian decode and native-int16 conversion.