cuda.core: forward ofast_compile to the NVVM backend once libNVVM supports -Ofast-compile #3206

Follow-up to #2954, which rejects ofast_compile on the NVVM backend together with the other NVRTC-only options. This issue tracks forwarding it instead, once libNVVM supports the flag.

What is known (2026-10-23):

• The libNVVM API documentation lists no -Ofast-compile option.
• On libNVVM from CUDA 13.4 (nvvm.version() 2.0, IR 2.0.3.2), compiling the textual IR from [BUG]: NVVM backend silently ignores 12 ProgramOptions fields that NVRTC emits #2954 with -Ofast-compile=0 succeeds; =min, =mid, and =max return NVVM_ERROR_COMPILATION with the log "parse Can't read textual IR with a Context that discards named Values"; =bogus returns NVVM_ERROR_INVALID_OPTION with "only supports 0, min, mid, or max". So the library knows the flag, and the failure is in the textual-IR parse path.
• A report in [BUG]: NVVM backend silently ignores 12 ProgramOptions fields that NVRTC emits #2954 on CUDA 13.2 (Thor) had all four levels accepted, for textual IR and for bitcode.

Asks:

1. Check whether -Ofast-compile works for bitcode input on 13.4, and whether the textual-IR failure is a regression.
2. Once libNVVM documents the flag, forward ofast_compile on NVVM under the same gate as NVRTC (is not None), for the inputs where it works. Replacing today's rejection with a pass-through is not a breaking change.
3. If the textual-IR failure is a libNVVM defect, file it against libNVVM and link it here.