[TableGen] Resolve arguments with fields in records

[optimisan]

Previously, instantiating an anonymous record happened in two steps:

1. Resolve the arguments with a MapResolver
2. Resolve the fields with a RecordResolver.

In 1, the MapResolver cannot resolve references to fields and in 2 the RecordResolver cannot resolve references to arguments. This was leading to incomplete resolution in llvm/test/TableGen/record-recursion.td (simplified test-case from #91050 ) like so:

1. MapResolver fails to resolve the if condition since it references the field v.
2. Resolving SumTillN<!sub(n, 1)>.ret) fails as it references the argument n and the RecordResolver doesn't know about it.

This patch merges the two steps into one where arguments and fields are resolved by the RecordResolver itself.

[arsenm]

Is it possible to just make this a required field?

[optimisan]

Yes, that would be ideal since the two-step resolution happens everywhere (in def R : C<n> as well).

This limitation of inaccessible arguments has forced the !if resolution to forego short-circuiting and resolve the MHS and RHS regardless of the condition. And so, the following case fails with a stack overflow:

class C<int n> {
    int v = n; 
    // !if ignores v==0 and evaluates C<n-1>.ret everytime
    // Ideally should short-circuit to 3
    int ret = !if(!eq(v, 0), 3, C<!sub(n, 1)>.ret);
}
def R : C<0>;

I think Record should contain arguments from each sub-class along with their fields as usual. Will see how that goes. With that, this patch would not be necessary since Record itself would have access to its arguments.