Add optimal graph coloring to minimize shared FIFO buffer allocations

[potatoeeh]

The CMSIS-Stream scheduler uses nx.coloring.greedy_color() to assign shared buffers to FIFOs via graph coloring. This greedy heuristic can use more colors (buffers) than the graph's chromatic number.

This adds an exact backtracking-based coloring algorithm (optimal_coloring) that finds the minimum number of colors. Because the solver is expensive on large graphs, it includes a configurable visit budget (max_visits, default 131072). If the budget is exceeded, the function returns None, allowing the caller to fall back to the existing greedy algorithm.

Two-tier coloring strategy:

1. Try exact backtracking (optimal, but potentially expensive)
2. If budget exceeded, fall back to greedy (suboptimal, but fast)

Also adds Tests/test_optimal_coloring.py exercising correctness, greedy-vs-optimal gap, and budget fallback behavior.

[christophe0606]

@potatoeeh Thanks for this very interesting PR. I'll be slow to merge it because I have currently lots of ongoing other tasks. So be very patient but I'll look at it and merge it.

Memory optimization in CMSIS Stream is something that can be improved with more work. There are a few places in the code where I could relax some constraints or use better algorithm so definitely this PR is appreciated.

Thanks.

I am curious to know in what kind of applications you use CMSIS-Stream ?

[christophe0606]

Merged. I'll update the package on PyPI soon.