[CUDA][PTXAS] Make TRITON_PTXAS_BLACKWELL_PATH work to overwrite ptxas-blackwell binary (#8945)

Mainly for THOR as by default ptxas-blackwell (cuda 12.9) would be used
and it would go "I don't understand sm110a".

Context: There is an environment variable TRITON_PTXAS_BLACKWELL_PATH
that was designed for users to overwrite this ptxas-blackwell, via the
lines noted in
#8945 (review).
But: `export TRITON_PTXAS-BLACKWELL_PATH=/usr/local/cuda/bin/ptxas` is
NOT a valid bash script, but
`export TRITON_PTXAS_BLACKWELL_PATH` is. So this PR maps
`ptxas-blackwell` binary to look for `TRITON_PTXAS_BLACKWELL_PATH`. IF
set, the ptxas-blackwell would use the provided "path" (nit: it is call
a "path" but we should supply the binary, like
"TRITON_PTXAS_BLACKWELL_PATH_AND_BINARY".

Author: nWEIdia

python/triton/knobs.py

@@ -196,7 +196,8 @@ def __init__(self, binary: str) -> None:
         binary += sysconfig.get_config_var("EXE")
         self.binary = binary
         self.default_path = os.path.join(os.path.dirname(__file__), "backends", "nvidia", "bin", binary)
-        super().__init__(f"TRITON_{binary.upper()}_PATH")
+        # Convert ptxas-blackwell to PTXAS_BLACKWELL, not PTXAS-BLACKWELL
+        super().__init__(f"TRITON_{binary.upper().replace('-', '_')}_PATH")
 
     def get(self) -> NvidiaTool:
         return self.transform(getenv(self.key))