Use correct features padding for encoder input

[MahmoudAshraf97]

This adheres faster whisper implementation to the original OpenAI implementation as discussed in #1084

These are WER comparisons before and after
These figures can be reproduced by running benchmarks/yt_commons.py and switching the batched inference to sequential

word_timestamps=False,
without_timestamps=True,
vad_filter=True,

Model	Before WER	After WER
distil-large-v3	26.277	14.762
distil-large-v2	71.848	81.456
distil-medium.en	68.044	66.565
distil-small.en	68.719	67.220

the performance regression of distil-large-v2 can be ignored because distil-large-v3 should be used as a drop-in replacement

This should also affect all use cases where chunk length is less than 30 for all models
There is also an average WER improvement by around 2% (relative) for batched inference across all models

Model	Before WER	After WER
tiny.en	15.437	15.063
tiny	21.765	21.390
base.en	14.300	13.816
base	17.709	17.251
small.en	13.054	12.617
small	16.413	16.088
medium.en	13.299	12.894
medium	15.991	15.593
large-v1	19.458	19.590
large-v2	15.237	15.148
large-v3	16.514	15.997
large-v3-turbo	14.576	14.044
distil-small.en	14.004	13.918
distil-medium.en	14.074	13.972
distil-large-v2	13.419	13.574
distil-large-v3	13.688	13.533

[MahmoudAshraf97]

This PR disabled the ability to change the encoder input and output dimensions (audio_ctx), any chunk_length passed to the transcribe will be respected, but the encoder input will be padded to 30s equivalent regardless of the chunk length

#171 (comment)