Render mutant copies by splicing text instead of rebuilding trees

[sknigh]

In my testing, the mutant generation phase ran about 3x faster with text splicing. There is a small compatibility risk because this implementation relies on an internal libCST API, which won't have the same compatibility SLA as public facing parts. On the other hand, this internal class is small (25 lines), has been stable for years, and a lot of libCST is built on top of it. As explained later, the tree build method remains available.

---

A mutant is a copy of its function with one node replaced. Generation rebuilt the function tree (deep_replace) and rendered it once per mutant.

Now the function is rendered once through a CodegenState subclass that records the output span of its name and of every mutation target, with the indentation in effect there. A mutant's text is the function's text with those two spans replaced; the replacement node is rendered with the recorded indentation, so the result matches rendering the rebuilt tree. A PrerenderedFunction node carries the text through module assembly and emits it verbatim, so line spans and the rest of the pipeline are unchanged.

A mutant whose target was rendered more than once, or precedes the function name, falls back to the tree rebuild. CodegenState is libcst internal API, but libcst's own node classes are built on it.

The test generates the generation corpus, plus sources with nested blocks, decorators, multi-line expressions, tabs and continuation lines, both ways and compares code, mutant names, line spans and hashes.