python: remove direct dependency on numpy

[reyammer]

Fixes #1268.

There don't seem to be performance regressions. The mean inference time and other metrics are relatively noisy, but there doesn't seem to be a big change. Cherrypicked examples below.

Stats with numpy (old version):

Inference stats over 770 files (repeat=10):
  Min: 0.0730 ms
  Max: 12.0421 ms
  Mean: 5.2871 ms
  Median: 5.2845 ms
  Total: 4071.0329 ms

Stats without numpy (this new version):

Inference stats over 770 files (repeat=10):
  Min: 0.0687 ms
  Max: 10.6079 ms
  Mean: 5.2051 ms
  Median: 5.2412 ms
  Total: 4007.9172 ms

[gemini-code-assist]

Code Review

This pull request successfully removes the direct dependency on numpy, which is a great simplification. The changes in pyproject.toml and the removal of numpy imports are correct. The logic to replace numpy functions with pure Python equivalents in magika.py is also sound. I have one suggestion to improve the readability and maintainability of a loop that was refactored.