Always use discriminant in ZeroTrie Serde; use serialize_bytes consistently

[sffc]

Fixes #4562

There are diffs; good to get this into 1.5

[sffc]

I scope-creeped a little bit and added more rmp-serde tests. I eyeballed all the custom Serialize impls in the project and added some tests on ones that looked important or suspicious. I found and fixed 1 other instance where we were using serialize_seq with deserialize_bytes.

[sffc]

Note: if you don't want me to serialize the whole thing as bytes, one potential alternative is that I could serialize a single u8 followed by bytes.

[sffc]

Note: if you don't want me to serialize the whole thing as bytes, one potential alternative is that I could serialize a single u8 followed by bytes.

^ Do you think doing it this way, with a separate entry encoding the discriminant, might be superior?

[robertbastian]

Yes I think it would be.

[sffc]

Ok coming right up

[sffc]

ok I have it implemented, but there's a stable key time_zone/iana_to_bcp47@1 that uses ZeroTrie and therefore the old format. I want to get rid of ZeroTrie and stick with just the variants (#4408) so I'm inclined to still use the new serialization format for the variants if we think it's superior.