Faster loops over lists

Floati's new just-in-time compiler learned the single most common shape of number-crunching on the real web: loops that walk an array by index -> adding up a column of numbers, doubling every value, transforming a buffer of pixels. Last release the fast path only handled loops over plain counters and variables; now it also compiles the array reads and writes inside them, so this whole class of work runs on the fast path instead of the slower interpreter. In our benchmarks a loop that sums a million-element array is about twice as quick, and one that rewrites every element in place is nearly twice as quick. It stays completely safe: the fast path only takes over for real dense arrays with whole-number positions, and the instant a loop touches anything it does not handle -> a gap in the array, an out-of-range position, a non-number, anything unusual -> Floati steps straight back to the interpreter with the results unchanged. Every array write is held back until the step it belongs to fully succeeds, so a step that falls back is redone once, never twice. The full engine test suite passes identically with the speed-up switched on and off.