[TensorIR][UX] Type annotation-based runtime type checking

[junrushao]

Addressed the UX issue raised in #9551.

To address the first issue (cryptic error message on packed function boundary), this PR introduced a decorator @type_checked to do runtime type checking and generate more user-friendly error messages. In our example:

The second issue is mainly wording. This PR rephrasing the error message to:

ValueError: The loop no long exists in the IRModule

However, I cannot find any environment variable to disable stacktrace printing. It was there months ago but no longer exists...Weird...

The notebook: https://gist.github.com/junrushao1994/cd0a33dc966da3b3f1abfa50a14e6d2c

CC: @tqchen

[junrushao]

Looks like we need at least python 3.7 to run the decorator…

[tqchen]

on the c++ stacktrace side, perhaps the main thing is to update the trace normalizer to simplify and skip some of the less informative traces(e.g. unpackcall_dispatcher) but keep the rest of the useful traces

[junrushao]

on the c++ stacktrace side, perhaps the main thing is to update the trace normalizer to simplify and skip some of the less informative traces(e.g. unpackcall_dispatcher) but keep the rest of the useful traces

Alternatively, there are other metaprogramming tricks to completely avoid recursion and could potentially generate significantly better error message.

[junrushao]

Update with a py36 compatible implementation

[junrushao]

The error message is improved to:

TypeError                                 Traceback (most recent call last)
/tmp/ipykernel_307968/1279557970.py in <module>
      5 i = sch.get_loops(block_c)
      6 # Tile the loop nesting.
----> 7 i_0, i_1, i_2 = sch.split(i, factors=[None, 4, 4])
      8 print(sch.mod.script())

~/Projects/tvm-dev/python/tvm/tir/schedule/_type_checker.py in wrap(*args, **kwargs)
    147                 )
    148                 if error_msg is not None:
--> 149                     raise TypeError(error_msg)
    150         return func(*args, **kwargs)
    151 

TypeError: In "Schedule.split", "loop" has wrong type. Expected "LoopRV", but gets: "list"