Fix issues in @errorCast with error unions

[Vexu]

Closes #17353
Closes #17355

#17354 passes analysis but crashes in codegen because error_set_has_value doesn't work with adhoc inferred error sets and I'm not quite sure how it should be fixed.

[andrewrk]

non-adhoc inferred error sets are intended to be resolved at the end of function body analysis:

zig/src/Module.zig

Lines 4896 to 4915 in 412d863

	// Resolving inferred error sets is done *before* setting the function
	// state to success, so that "unable to resolve inferred error set" errors
	// can be emitted here.
	if (sema.fn_ret_ty_ies) |ies| {
	sema.resolveInferredErrorSetPtr(&inner_block, LazySrcLoc.nodeOffset(0), ies) catch |err| switch (err) {
	error.NeededSourceLocation => unreachable,
	error.GenericPoison => unreachable,
	error.ComptimeReturn => unreachable,
	error.ComptimeBreak => unreachable,
	error.AnalysisFail => {
	// In this case our function depends on a type that had a compile error.
	// We should not try to lower this function.
	decl.analysis = .dependency_failure;
	return error.AnalysisFail;
	},
	else => |e| return e,
	};
	assert(ies.resolved != .none);
	ip.funcIesResolved(func_index).* = ies.resolved;
	}

adhoc inferred error sets are only used for inline/comptime calls and are resolved after the call:

https://github.com/ziglang/zig/blob/412d863ba5801c1376af7ab8f04a71b839a820a6/src/Sema.zig#L7542-L7553

In both cases, they should not be observable by the backend.

[andrewrk]

The logic in this block should be skipped for error unions.

[Vexu]

I use it to avoid an error_set_has_value in the safety check but it could be removed too.

[andrewrk]

Ah, my mistake