sema: coerce comptime known arguments earlier for @min @max

[IOKG04]

Currently an assertion fails when trying to compile code like the following.

const a = @max(@as(f32, 1.0), std.math.maxInt(u32));
// or
var f: f32 = 1.0;
_ = &f;
const b = @min(f, std.math.maxInt(u32));

As far as I can see, this happens because in sema.analyzeMinMax no source location is given to the coercion of std.math.maxInt(u32) to the result type (f32). Because std.math.maxInt(u32) isn't representable as an f32 without rounding, this will throw an error. The error code then asserts that there is a source location, which, as said before, there is not.

I fixed this by coercing the comptime known arguments a bit earlier, with a source location given.
Also I removed the coercion of the comptime known argument(s) to intermediate_scalar_ty later on, because they are already guarantied to be coerced into it.

fixes #25284

[IOKG04]

I just noticed that I forgot to add tests so this doesn't regress. Will do when I'm back from school.

[IOKG04]

I wrote in one of the commit messages that running the tests crashed my pc.
As it turns out, that was cause .zig-cache took up ~a quarter tebbibyte, over-filling my drive..
Based on tests passing, I'd assume I just had an unlucky case with some weird bug.
Also heard from a friend that this is a kind of known issue, except it can't be replicated..

Well in case it helps with figuring it out, I'm using Arch Linux with a Ryzen 7 5800x, 32 GiBs of RAM and a 960 GiBs NVME SSD as my /home/ drive.

#25192 seems related.