Convert remaining NaN tests to use new assert types

[Cellule]

Use assert_return_canonical_nan and assert_return_arithmetic_nan instead of assert_return with a nan expected value.
It has been a while since those should have changed to assert_return_nan, but now that we've introduced 2 new types of asserts for NaN, I thought it would be a good time to clean up.

[Cellule]

It seems the interpreter considers the value 0x7fa00000 to be an invalid type of nan in the test
(assert_return_arithmetic_nan (invoke "f32.reinterpret_i32" (i32.const 0x7fa00000))) (conversion.wast)

With the implementation

| [Values.F32 got_f32] -> let pos_nan = F32.to_bits F32.pos_nan in
        Int32.logand (F32.to_bits got_f32) pos_nan = pos_nan

0x7fa00000 & 0x7fc00000 != 0x7fc00000

I read WebAssembly/design#660 one more time, but I can't figure out what is the correct thing to do in this case.
Also, I am not very fluent in caml and could use some help on this

[Cellule]

I think I found the problem, I think we should use bare_nan instead of pos_nan to check if it's a valid nan for assert_return_arithmetic_nan

[rossberg]

Actually, various uses of assert_return on a NaN are intentional: they specifically test for exact bit patterns. Wasm is only loose on the bit patterns of NaNs generated from arithmetic operations, in other cases like reinterpret or memory reads/writes the exact bit pattern must be preserved. The new NaN assertions are special cases to be used only where the result value is underspecified, and multiple values are possible.

[sunfishcode]

The promote/demote test changes are right; it looks like we missed those when we converted to the new asserts. The others are as @rossberg-chromium says. In particular, abs, neg, and copysign are also defined to be bitpattern-preserving (in accordance with IEEE 754).