Performance improvement opportunities in `compute_finder_penalty_score`

[anforowicz]

Let me open an issue to capture my notes about performance improvement opportunities in compute_finder_penalty_score. Frequent boxing should be taken care of by 29296f0, but there may be additional improvement opportunities:

• It seems that this function implements a substring search algorithm. The current implementation uses naive string searchwhich works in O(N x M) time, where N is the size of the "hay" (e.g. the width of the QR code) and M is the size of the "needle" (e.g. the size of the undesirable locator pattern - always 7 items in length). O(N + M) time should be possible using KMP or Rabin-Karp algorithm.
• This function extracts rows/columns with the indirection of a get lambda. Maybe the indirection is optimized away by the compiler, but maybe using subslicing of Canvas::modules would be more efficient (since rows are just continuous slices). Avoiding this indirection may also be applicable to compute_adjacent_penalty_score and compute_block_penalty_score.

I don't plan to work on this, but I hope that leaving the notes above will help the next person who might want to pick it up.