Validate natural alignment for atomic memory operations

[ANAMASGARD]

Fixes #8962

Atomic memory instructions were parsing memarg alignment but not keeping it in the IR for RMW, cmpxchg, wait, and notify. The builder also forced natural alignment, so invalid alignment from WAT or binary input was silently normalized and passed validation.

The threads proposal requires atomic accesses to use natural alignment (see WebAssembly/threads#242). Binaryen should reject invalid modules rather than accept them.

This PR stores align on all atomic memory IR nodes, threads it through WAT/binary parsing and printing, and validates it centrally via the existing validateAlignment(..., isAtomic=true) check. The DeAlign pass skips atomic load/store so it does not rewrite their alignment. Adds lit tests and unskips threads/atomic.wast.

Test plan

• test/lit/validation/atomic-alignment.wast
• test/lit/passes/dealign-atomics.wast
• threads/atomic.wast (unskipped)

[stevenfontanella]

Thanks for the PR! Will take a look in the next few days

[stevenfontanella]

Thanks for the PR!

One high-level comment on the approach: this PR stores a redundant align field in the AST for each AtomicRMW, AtomicCmpxchg, and atomic wait/notify expression. e.g. for a 32-bit atomic RMW, the align is always 4 and anything else would be invalid, so there's really no need to store this except for the validation check. For load and store on the other hand, it does need to be stored because the non-atomic versions of these could have any alignment and we need to preserve that in the binary.

So let's change the approach and make this a parse-time failure instead. This is different from what the spec says but in general we don't distinguish the two in Binaryen (the spec test runner allows either type of failure for assert_invalid and assert_malformed). This way we don't need to store the alignment in the AST for these expressions.

[stevenfontanella]

Looks like there are some existing tests that don't use a natural alignment e.g. here and here. We can change them to load from address 0.