Enable JIT test for METAL. Add METAL_NO_FAST_MATH flag.

[chenyuxyz]

Metal by default enables fast-math, which may violate IEEE 754 standard (https://developer.apple.com/documentation/metal/mtlcompileoptions/1515914-fastmathenabled). With fast-math, rand * 2 * pi (used in randn implementation) will be different from numpy about 30% of the time, which breaks the exact number tests in test_jit.py. test_ops.py was not affected because it's all testing allclose

Add a flag METAL_NO_FAST_MATH to disable fast-math.

[tinyb0t]

Changes made in tinygrad/:

------------------------------------------------------------
files                             insertions       deletions
------------------------------------------------------------
tinygrad/jit.py                            3               1
tinygrad/runtime/ops_metal.py              3               0
------------------------------------------------------------
total                                      6               1
------------------------------------------------------------
lines added in the tinygrad folder: 5

[geohot]

Tests failing, but overall I like the change

[chenyuxyz]

METAL test behaves differently on my local M1 vs CI. Might be related to other Metal issue. WebGPU fails because it's using METAL backend. I marked this as draft now unless we want to merge one more skipped test for METAL?

[chenyuxyz]

I have a better way to improve this and I want to separate (1) enable JIT test in CI for all backends, (2) improve JIT tests (currently some tests still pass if we remove TinyJIT decorator), and (3) flags to control metal compile.

I will close this one and open new ones separately to make review easier.