Allow non-canonical type alignment?

Over in wasi-nn, we have been wondering whether there might be a way to specify an alignment for a type larger than the one specified by the canonical ABI. This is similar to what @titzer brings up in https://github.com/WebAssembly/component-model/issues/194 but I thought it was different enough to merit its own issue.

wasi-nn currently specifies the tensor data as list<u8> (WIT). The initial rationale (long ago, back in the WITX days) was a performance consideration: ML frameworks can deal with tensor data as byte blobs and if we can just "pass the bytes through" we can avoid some extra copies. This means that when we, as a wasi-nn guest user, receive tensor data back we must convert it somehow from list<u8> to list<f64>, e.g. This conversion is an unsafe cast in some of the bindings — what if the list<u8> is not aligned to cast to list<f64>? There has been an outstanding issue (https://github.com/WebAssembly/wasi-nn/issues/27) to figure this out. One solution would be to always copy the data; another to conditionally copy the data only if unaligned; etc.

The solution I'm asking about here is adding some way to inform the runtime of the alignment we would like when allocating space for the tensor data. Is this possible? Is it different enough from what @titzer is asking? I suspect we would only want to allow allocations for alignments larger than what the canonical ABI specifies, but maybe not.

Start by reading the linked component-model issue #194, wasi-nn's wit/wasi-nn.wit at line 44, and wasi-nn issue #27 to understand the canonical ABI alignment and tensor conversion concerns. Done means reaching and documenting a decided design for whether callers can request larger alignments and how that relates to the existing proposal.