perf(inlet): avoid redundant list rebuild in pull_sample

[sappelhoff]

(PR and content generated with the help of Claude)

What

Avoid a redundant list rebuild in StreamInlet.pull_sample without changing its output.

Why

The old code always built an intermediate list with a comprehension, then for string streams re-iterated it a second time to decode:

sample = [v for v in self.sample]
if self.channel_format == cf_string:
    sample = [v.decode("utf-8") for v in sample]

Now the numeric path uses list(self.sample) (faster than a comprehension), and the string path decodes directly from the ctypes sample in a single pass with one fewer allocation:

if self.channel_format == cf_string:
    sample = [v.decode("utf-8") for v in self.sample]
else:
    sample = list(self.sample)

Impact

Measured on the extraction step alone (output verified identical via assert old() == new()):

channels	numeric old/new	string old/new
8	0.69 / 0.62 us (1.12x)	1.33 / 1.17 us (1.14x)
32	2.02 / 1.77 us (1.14x)	4.23 / 3.89 us (1.09x)
64	3.88 / 3.39 us (1.15x)	8.23 / 7.51 us (1.10x)

A modest ~1.1-1.2x per call; negligible at 1 channel. pull_sample is the most-called function in per-sample loops, so it adds up there, and the string path also drops one throwaway allocation per call.

Behavior

Identical output: same list of values, same string decoding, unchanged legacy assign_to and (None, None) paths. Verified with a live localhost round-trip (numeric, string, legacy assign_to); existing test suite passes.

[cboulay]

LGTM. Thanks!