Emit valid kernel source for non-finite float constants - #3648

Compiling a function that bakes a NaN (or infinity) scalar constant into a
fused kernel failed because print_float_constant streamed a bare token such
as `nan`, which is not a valid identifier in the generated kernel source
(e.g. Metal: "use of undeclared identifier 'nan'").

Emit the NAN / INFINITY / -INFINITY macros instead; both the Metal
(<metal_stdlib>) and CPU (<cmath>) kernel toolchains provide them. The
finite check widens to double first so it is exact for every float type.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

The prebuilt CPU preamble is produced by running the preprocessor over
compiled_preamble.h, which consumes the NAN / INFINITY macro definitions
(macros vanish during preprocessing, leaving only the expanded <cmath>
declarations). print_float_constant emits those macros for non-finite
constants, so the generated kernel failed to compile with "'NAN' was not
declared in this scope" on the Linux/Windows prebuilt-preamble path.

Define the macros in the generated source when they are missing, backed
by std::numeric_limits (available via <limits>). The #ifndef guards make
this a no-op on paths where <cmath>'s macros are already in scope (e.g.
the include-based preamble path), avoiding redefinition clashes.

Also drop the now-dead `#define inf` from the CUDA backends: non-finite
constants are emitted as the standard INFINITY/NAN macros (which CUDA
provides), so the bare `inf` token is no longer generated.

Verified on a Linux + g++ build (the failing path): the regression test
fails with the original "'NAN' was not declared" error without this
change and passes with it.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>