Prevent data copy in VideoFrame.to_ndarray() for padded frames

[lgeiger]

Calling VideoFrame.to_ndarray() on unaligned video frames is very slow since the frame is copied.

This PR prevents the copy for padded frames by using strided views into the frame data.

When benchmarked with the following code I'm seeing a huge speedup for padded frames. Which now has the same performance characteristics as the aligned case

import timeit
import numpy as np
import av

aligned = av.VideoFrame(1920, 1080, format="rgb24")
padded = av.VideoFrame(1080, 1920, format="rgb24")
aligned_yuv = av.VideoFrame(1920, 1080, format="yuv420p")
padded_yuv = av.VideoFrame(1080, 1920, format="yuv420p")

n = 1000
t_aligned = timeit.repeat(lambda: aligned.to_ndarray(), repeat=5, number=n)
t_padded = timeit.repeat(lambda: padded.to_ndarray(), repeat=5, number=n)
t_aligned_yuv = timeit.repeat(lambda: aligned_yuv.to_ndarray(format="rgb24"), repeat=5, number=n)
t_padded_yuv = timeit.repeat(lambda: padded_yuv.to_ndarray(format="rgb24"), repeat=5, number=n)

t_aligned = np.array(t_aligned) / n
t_padded = np.array(t_padded) / n
t_aligned_yuv = np.array(t_aligned_yuv) / n
t_padded_yuv = np.array(t_padded_yuv) / n

print(f"Aligned (1920x1080 rgb24):   {t_aligned.mean() * 1e6:.2f} ± {t_aligned.std() * 1e6:.2f} µs")
print(f"Padded  (1080x1920 rgb24):   {t_padded.mean() * 1e6:.2f} ± {t_padded.std() * 1e6:.2f} µs")
print(f"Aligned (1920x1080 yuv420p): {t_aligned_yuv.mean() * 1e6:.2f} ± {t_aligned_yuv.std() * 1e6:.2f} µs")
print(f"Padded  (1080x1920 yuv420p): {t_padded_yuv.mean() * 1e6:.2f} ± {t_padded_yuv.std() * 1e6:.2f} µs")

Before:

Aligned (1920x1080 rgb24):   1.76 ± 0.02 µs
Padded  (1080x1920 rgb24):   148.09 ± 7.51 µs
Aligned (1920x1080 yuv420p): 364.42 ± 1.05 µs
Padded  (1080x1920 yuv420p): 522.63 ± 0.08 µs

After:

Aligned (1920x1080 rgb24):   1.73 ± 0.03 µs
Padded  (1080x1920 rgb24):   1.76 ± 0.04 µs
Aligned (1920x1080 yuv420p): 365.67 ± 0.83 µs
Padded  (1080x1920 yuv420p): 368.43 ± 0.86 µs

The main changes of this PR are in e2ad2fb.

b105f6e, cd5ffd0, and 02ec746 are minor followup optimisations to ensure that this PR doesn't introduce a regression for the aligned case.