Honour 'pragma: no mutate block' on else/except/finally arms

[dylanpulver]

_visit_compound_header is only reachable from visit_If / For / While / With / Try / FunctionDef / ClassDef / Match. In libcst, else, except, except* and finally are separate nodes owning their own suite, so a # pragma: no mutate block on one of those header lines reaches no visitor and is dropped — silently, which is the bad part: you mark an error branch as skipped and it gets mutated anyway.

Measured on main, ignore_node_lines for a pragma on each header:

if       [2,3]   works        else (if/for/while/try)   []   ignored
elif     [4,5]   works        except                    []   ignored
while    [2,3]   works        except*                   []   ignored
with     [2,3]   works        finally                   []   ignored
try      [2,3]   works        try  (of a try/except*)   []   ignored

elif works because it parses as a nested If; the else one line below it does not. The last row is the same source line behaving differently depending on whether a later handler writes except or except* — try/except* parses as TryStar, and there is no visit_TryStar.

README:291 says the block pragma "works on any compound statement -- functions, classes, if/elif/else, loops, context managers, etc."; README:345 extends it to for/else, while/else and try/except/finally.

Five visitors added, in the existing style. Match/MatchCase deliberately untouched — #554 is fixing that, and I did not want to duplicate it. I cherry-picked this commit onto pull/554/head: it auto-merges with zero conflicts and the combined tree passes 50/50 in test_pragma_handling.py. Land them in either order.

Checked by reverting: with pragma_handling.py back at main and the tests kept, exactly the 8 new tests fail and the other 36 pass. Adding only visit_Else still fails 4 of them (except, except*, finally, TryStar), so each visitor is pinned rather than carried by its neighbours. A sibling-leak case asserts the statement after the else block stays mutable.

uv run pytest gives 387 passed / 2 failed against 379 passed / 2 failed on main, same venv after uv sync (three consecutive runs each). The two failures are the test_e2e_type_checking snapshot tests and are identical on both trees. test_safe_setproctitle is flaky here on macOS — it failed in one earlier run of each tree and in none of the three above. ruff check and ruff format clean at the pinned 0.15.7; mypy reports 19 errors, all in __main__.py, none in the files touched, identical to main. I did not run ./scripts/run_tests.sh in the Linux container.

Found by enumerating the visit_* methods against the libcst node types that own a suite, then probing each. AI assistance: this change and its tests were drafted with Claude Opus 5 (claude-opus-5). The commands and numbers above were run locally.

[nicklafleur]

one very minor nit otherwise lgtm