speed up the encoder kde by not calling np.histogram on the hot path

[jl33-ai]

Makes the per spike encoder kde about 3.3x faster: median goes from 2096 to 635 microseconds at a full 50k mark buffer.

• The cost was np.histogram. Given explicit bin edges it does a binary search per sample. Our position bins are evenly spaced integers, so a direct index plus np.bincount does the same job much faster.
• Also preallocates the kde scratch instead of allocating about a dozen temporary arrays per spike, and fuses the squared distance into a single np.einsum pass.
• Same result to within floating point rounding, about 3e-13. Set encoder.mark_kernel.exact_histogram to true to restore the exact original numerics.
• Verified by benchmarks/bench_hotpath.py --verify, which matches the old code on every spike of a 400 spike run.
• The decoder posterior was measured too (65 to 173 us per bin) and left alone, it is not a bottleneck.
• Adds the benchmark and docs/latency_analysis.md, which shows end to end latency is dominated by the time bin width and the delay_samples jitter buffer, not the compute.
• Heads up: the fast path is on by default, so output shifts at the 1e-13 level. Happy to make it opt in instead if you prefer.