[Leitwolf11]

Starts the frontend RTS with -A64m -n4m instead of the default 4 MiB nursery.

The frontend grows linearly with the program, so there was no quadratic path to fix, and GC was about a third of its time. The AST stays live, so the small nursery kept collecting and copying it.

Six alternating runs on a generated program of 13 sources and about 4800 methods, optimized talac, --emit=core, on a noisy machine:

Options	Min	Median	GC median	Total memory
default	4.98 s	7.58 s	3.05 s	346 MiB
-A64m -n4m	4.13 s	4.61 s	1.60 s	331 MiB

-A64m without -n4m showed no clear gain in single runs, and -O512m added nothing. tala-compiler tests pass on Windows, and the generated program builds and runs. tali shares this library, so it gets the same nursery.

Self-review: rts_opts is the same channel as -with-rtsopts, so it needs no rts_opts_enabled change and users still cannot pass RTS flags. Rts.h replaces HsFFI.h, which it includes, and is needed because RtsAPI.h does not stand alone. It prints deprecated-volatile warnings from GHC's own headers in C++20, which do not fail the build. Linux and macOS are left to CI.

[Leitwolf11]

Self-review of the later commit: renameWhenReleased retries only on fs.ErrPermission, for at most ten seconds, so any other failure still surfaces at once. The Windows job passed on this head, which had failed twice on that rename before.