PERF: avoid unnecessary array copy for single-pass convolution

[neutrinoceros]

Close #85
Incidentally, close #114
I checked that binary size doesn't noticibly increase with this refactor, lifting my comment from #85

[neutrinoceros]

This is stable. Before I undraft, squash and merge, I want to address two points:

• document bound memory usage as a feature (update: this is DOC: document memory usage as a feature (peak and output) #175)
• check if this impacts performance for single pass. A comment I left on PERF: avoiding an array copy for single-pass convolution #85 suggests that it should moderately reduce run time, but it's worth checking.

[neutrinoceros]

I'm actually seeing a 5% performance regression with this patch. I'll re-issue the various independent parts of the PR piecemeal to reduce my cognitive overload deciphering this.

[neutrinoceros]

I think I've been looking at this from the wrong angle; I was assuming that my convolve_once was somehow less efficient than convolve_iteratively, while in fact, I'm seeing 5% overhead with both implementations, so maybe the dispatching itself is generating overhead (pretty surprising), or there's something else I'm missing.

[neutrinoceros]

(dispatching at the Python level instead doesn't improve performance)

[neutrinoceros]

(forcing inlining on convolve_once and convolve_iteratively doesn't help either)

[neutrinoceros]

The good news is that #172 more than compensates for the loss in performance from this patch, but I still would like to figure out why this is slower than main.

[neutrinoceros]

rebased, but I'm still seeing about 2% (unexplained) overhead.