[TIRScript] fix parse StringImm value in for loop annotations

[wrongtest-intellif]

Fix a minor issue that for i in T.serial(0, 4, annotations={"pragma_key":"value"}) result to a runtime.String typed dict value. The flatten pass do not allow attribute value of runtime.String .

[vinx13]

Annotations actually expect String as the key, can we fix the pass instead?

[wrongtest-intellif]

@vinx13 Hi, the issue is about string dict value as a clarification. The key type seems no problem. The parser already convert str to StringImm in annotations but provide the old dict (maybe a typo).

For the pass, currently it assumes all dict values should be of PrimExpr in annotations. Did you mean we can allow some "compatible" values to convert to PrimExpr instead of simple downcasts?
https://github.com/apache/tvm/blob/main/src/tir/transforms/flatten_buffer.cc#L99-L106

[vinx13]

StringImm here might be something I forgot to change in #9742. Annotations value is ObjectRef and we would like to use String directly.
In FlattenBuffer, due to compatibility issue with legacy AttrStmt, we should handle String annotation value and create StringImm in that case

[wrongtest-intellif]

handle String annotation value and create StringImm

I update it by the suggestion. But another problem is detected. Primfunc with flattened loop attrs seems not supported by script parser. It would lead to "use before def" errors.

def main(A: T.Buffer[(16,), "float32"]) -> None:
        # body
        T.attr(i, "pragma_1", "str_value")
        ...
        for i in T.serial(16):

[junrushao]

@Hzfengsy is that exactly the same issue we talked about a while ago?

BTW please rebase to our latest HEAD

[Hzfengsy]

Thanks @wrongtest @vinx13 @junrushao1994