Allow unparenthesized walrus in set literals, set comprehensions and indexes

[gousaiyang]

This incorporates two recent changes in CPython upstream:

• As per bpo-42381: Allow walrus in set literals and set comprehensions python/cpython#23332, unparenthesized assignment expressions are allowed in set literals and set comprehensions (but not in dict literals and dict comprehensions). This change applies to Python 3.9+. Examples: {x := 1, 2, 3}, {x4 := x ** 5 for x in range(7)}.
• As per bpo-42316: Allow unparenthesized walrus operator in indexes python/cpython#23317, unparenthesized assignment expressions are allowed in indexes (but not in slices). This change applies to Python 3.10+. Examples: a[b:=0], a[b:=0, c:=0].

As for the code change, I cannot directly use namedexpr_test in dictorsetmaker and subscript because that will produce ambiguous rules.

[isidentical]

I believe it looks fine, thanks @gousaiyang! @davidhalter can you check this out? "The ambiguity arises in"

[davidhalter]

Sorry for the delay (was awaiting my firstborn :)), but thanks for the change!

I think this is definitely fine, but I will check why this causes ambiguities.

IMO we will have to find a (PEG) solution for the parser anyway during the next year, so this grammar change is not a problem anyway.

[isidentical]

Sorry for the delay (was awaiting my firstborn :)

🚀 Congratulations!

[davidhalter]

I just tried to check why the ambiguity problem arises:

subscript: test [':=' test] | [test] ':' [test] [sliceop]
vs.
subscript: namedexpr_test | [test] ':' [test] [sliceop]

The problem in the latter example is that for all the first tokens (e.g. for a lambda token), the LL parser does not know if it should use test or namedexpr_test, because there's an alternative [test].

So this is definitely a limitation of our current parser.

[davidhalter]

Also released 0.8.1 with this.