Handle non-contiguous input in process()

[sanerdemirel]

Fixes #3.

The problem

process() copied each channel's samples with std::copy under the assumption
that the input buffer was C-contiguous:

std::copy(inData + i*inputLength, inData + (i+1)*inputLength, inputChannels[i]);

A transposed array — which is what librosa.load(..., mono=False) commonly
hands you — is a view with the channel and sample strides swapped relative to a
freshly allocated one. Copying through it as if it were still contiguous reads
the wrong samples, so the output is silently corrupted rather than failing.

The change

process() now reads each sample through the array's own strides instead of
assuming a layout. The input was already being copied into internally-owned
buffers before processing, contiguous or not, so this costs nothing extra for
the contiguous case — it just also gives the right answer for the
non-contiguous one.

While in there, process() also accepts 1-D (mono) input by promoting it to a
single channel instead of raising TypeError, matching the layout handling
proposed on the issue.

How I checked it

• Contiguous input is bit-identical to before, across 6 configurations
  (mono and stereo, several time factors).
• With a positive control, because "bit-identical" is worthless if the
  comparison could not have detected a difference in the first place: I
  confirmed the harness could tell the old and new builds apart when a
  difference was deliberately introduced, before trusting the runs where it
  reported none.
• Non-contiguous input now matches contiguous input for the same audio.

tests/test_noncontiguous_input.py covers the transposed-input case in both
mono and stereo.

Note

Opened as a draft. Happy to adjust anything about the approach or the tests.