Optimize away overhead from setting up try catch if exception is impossible

[oxinabox]

There are a bunch of senarios, in which is should be possible to know at compile time if an exception will be thrown. (especially if we exclude unrecoverable externally triggged exceptions like out of memory).

E.g. if the code is inlined to only builtin's / LLVM intrinsics,
and there is no exception present in the code.
(maybe there could be for different iinput types, but we specialized on that so those go away).

Right now, even in the most obvious case, the compiler can't tell.
So it includes all this overhead to be ready to catch exception.

Example:

julia> function danger()
           1.0
       end
danger (generic function with 1 method)

julia> function safe()
           try
               1.0
           catch
               rethrow()
           end
       end
safe (generic function with 1 method)

julia> @btime safe()
  17.163 ns (0 allocations: 0 bytes)
1.0

julia> @btime danger()
  0.031 ns (0 allocations: 0 bytes)
1.0

Realistic case:
JuliaDiff/ChainRulesCore.jl#82

[oxinabox]

Closed by #51674

julia> @btime safe()
  0.855 ns (0 allocations: 0 bytes)
1.0

julia> @btime danger()
  0.874 ns (0 allocations: 0 bytes)
1.0