Fix AtWithRecord

[leezu]

Prior implementation recorded reshape with concrete shapes used during the particular invocation of At.
New implementation uses records reshape with magic numbers to match symbolic interface.

Example of failure without this patch

=================================== FAILURES ===================================
______________________ test_gpt2_incremental_states[ctx0] ______________________

ctx = cpu(0)

    def test_gpt2_incremental_states(ctx):
        with ctx:
            batch_size = 4
            sequence_length = 5
            inputs = mx.np.random.randint(0, 1000, (batch_size, sequence_length), ctx=ctx)
    
            cfg = GPT2Model.get_cfg()
            gpt2_model = GPT2Model.from_cfg(cfg)
            gpt2_model.initialize(ctx=ctx)
            gpt2_model.hybridize()
    
            one_time_hiddens, one_time_states = gpt2_model(
                inputs,
                gpt2_model.init_states(batch_size, ctx)
            )
    
            states = gpt2_model.init_states(batch_size, ctx)
            hiddens_l = []
            for i in range(sequence_length):
>               hiddens, states = gpt2_model(
                    inputs[:, i:i+1],
                    states
                )

tests/test_models_gpt2.py:96: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/util.py:298: in _with_np_shape
    return func(*args, **kwargs)
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/util.py:482: in _with_np_array
    return func(*args, **kwargs)
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/gluon/block.py:1429: in __call__
    return self._call_cached_op(x, *args)
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/util.py:298: in _with_np_shape
    return func(*args, **kwargs)
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/util.py:482: in _with_np_array
    return func(*args, **kwargs)
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/gluon/block.py:1131: in _call_cached_op
    out = self._cached_op(*cargs)
/opt/hostedtoolcache/Python/3.8.5/x64/lib/python3.8/site-packages/mxnet/_ctypes/ndarray.py:171: in __call__
    check_call(_LIB.MXInvokeCachedOp(
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

ret = -1

    def check_call(ret):
        """Check the return value of C API call.
    
        This function will raise an exception when an error occurs.
        Wrap every API call with this function.
    
        Parameters
        ----------
        ret : int
            return value from API calls.
        """
        if ret != 0:
>           raise get_last_ffi_error()
E           mxnet.base.MXNetError: MXNetError: Error in operator node_16288: [02:57:54] ../src/operator/numpy/np_matrix_op.cc:149: Check failed: src.Size() == dst->Size() (6144 vs. 0) : Cannot reshape array of size 6144 into shape [2,4,0,768]

Details on failure

Json export of the recorded graph contains the following segment. We can see the shape of the first forward pass was hardcoded by the AtWithRecord implementation.

    {
      "op": "slice",
      "name": "node_10",
      "attrs": {
        "begin": "0",
        "end": "1"
      },
      "inputs": [[5, 0, 0]]
    },
    {
      "op": "_np_reshape",
      "name": "node_11",
      "attrs": {"newshape": "[2,4,0,768]"},
      "inputs": [[14, 0, 0]]
}

[mxnet-bot]

Hey @leezu , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [clang, miscellaneous, centos-cpu, windows-gpu, edge, sanity, unix-gpu, website, unix-cpu, centos-gpu, windows-cpu]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[sxjscience]

LGTM

[leezu]

@mxnet-bot run ci [centos-cpu, centos-gpu, edge, miscellaneous, unix-gpu]

#19373

[mxnet-bot]

Jenkins CI successfully triggered : [edge, unix-gpu, centos-gpu, centos-cpu, miscellaneous]

[leezu]

@mxnet-bot run ci [centos-gpu, edge, miscellaneous, unix-gpu, unix-cpu]

#19373

[mxnet-bot]

Jenkins CI successfully triggered : [miscellaneous, edge, centos-gpu, unix-gpu, unix-cpu]

[leezu]

@mxnet-bot run ci [miscellaneous, unix-gpu, unix-cpu]

[mxnet-bot]

Jenkins CI successfully triggered : [miscellaneous, unix-cpu, unix-gpu]

[leezu]

@mxnet-bot run ci [miscellaneous, unix-gpu]

[mxnet-bot]

Jenkins CI successfully triggered : [miscellaneous, unix-gpu]

[leezu]

@mxnet-bot run ci [unix-gpu]

[mxnet-bot]

Jenkins CI successfully triggered : [unix-gpu]