[fuzz] Rework how fuzzer sidecar data works

The existing implementation is based around capturing all of the comment
data from the shader input and effectively doing a reinterpret cast to
the specific transformation options struct.

Given that many of these structs have dynamically sized containers
(vectors, maps, etc), it is very unlikely that this will produce a valid
state, because it requires the value from the number of elements bytes
to correctly match the number elements that fit in the number of
remaining bytes. It also has other significant limitations with regards
to data over/under runs, etc.

This is being replaced with a custom structured decoding/encoding
implementation inspired by FuzzedDataProvider from libFuzzer,
https://llvm.googlesource.com/libfuzzer/+/refs/heads/main/FuzzedDataProvider.h
that uses the comment data as an entropy source.

This change also lets us remove bytes::Decode, which in turn allows for
dropping a bunch of other machinery in utils/bytes.

Fixes: 559142627
Change-Id: Icfedc1babaafdd5dc60263216521a3a2701c2ff7
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/341015
Commit-Queue: dan sinclair <dsinclair@chromium.org>
Reviewed-by: dan sinclair <dsinclair@chromium.org>
Auto-Submit: Ryan Harrison <rharrison@chromium.org>