Tighten versionsuffix rules and add WAST tests

[lukewagner]

This PR makes the change suggested in #735 of always requiring a valid semver string (when there is a version at all).

In adding some WAST tests for versionsuffix, I noticed a few other variations from the spec and wasm-tools as captured in this PR in test/validation/version-suffix.wast. In particular: the grammar currently only allows a versionsuffix when the version in the name is "canonical" (e.g., a:b/c@1.2 is not "canonical" so it shouldn't be allowed to have a versionsuffix according to the current grammar). Also, the current text parser only parses the attributes in a fixed order, whereas the grammar parses them in any order. Happy to hear arguments if yall think we should change the spec on either of these though.

One last thing on the topic of versionsuffix: we have since added the attribute external-id, which is hyphenated, so it seems like we should name the attribute version-suffix instead. Happy to make this change in this PR too.

cc @chenyan2002 for thoughts on this too

(I'll be out for until next week, sorry in advance for slow reply.)

[chenyan2002]

the grammar currently only allows a versionsuffix when the version in the name is "canonical"

I think this is implemented in bytecodealliance/wasm-tools#2693

[chenyan2002]

During the transition period, there is this strange mixed use of canonical name and full version:

    (import "a:b/c@1" (versionsuffix ".2.3") (instance))
    (import "a:b/c@1.2.3" (instance))

I guess we don't need to say much in the spec, but in the implementation, we do need to take care of this, and report it as an error.

[lukewagner]

Agreed, I didn't add anything to the spec for this since presumably this case goes away over time.

[alexcrichton]

s/is only/is/ perhaps? (to clarify it's always required for canonversion)

[lukewagner]

The update refers to the "Canonical Interface Version" section which now says that the concatenation must split at the "right" point, as defined by the pseudo-algorithm in that section, so I think that will end up forcing a suffix when you have a canonversion (except when the canonversion is valid semver on its own).

[lukewagner]

Thanks for the review! This commit updates the spec/tests to now cover the special pre-release rules (which required adding a new grammar case to canonversion for <pre-release> and also has the interesting consequence of making a grammatically-valid canonversion become invalid only once you've seen the versionsuffix) and also reject empty versionsuffix. I also changed the text format grammar from an unordered <attribute>* to an ordered attributes ::= <implements>? <versionsuffix>? <externid>? for now (until someone complains and wants to generalize to an unordered list).

With those changes and consequent WAST updates, test/validation/version-suffix.wast now passes on dev wasmtime 👍. PTAL

Only last question: could we update versionsuffix in the text format to version-suffix (to match external-id) now or sometime before releasing canonical version names as a feature? It seems like a good idea for regularity.

[alexcrichton]

could we update versionsuffix in the text format to version-suffix

yes!

[lukewagner]

Awesome, renamed in the last commit.