fix dot(csr.T, dns)=dns can't be called on cpu and gpu

[XiaotaoChen]

Description

This PR is to fix a bug that cause the implement code of dot(csr.T, dns)=dns can't be called on CPU and GPU. This is 10 times faster than before when test dot(csr.T, dns)=dns with density=0.0001.

The reason for this bug is the function of DotForwardInferStorageType in dot-ink.h haven't map this situation that input stype=[csr, default], output stype=[default] and transpose_a=True to the implement of dot(csr.T, dns)=dns

Details

The following code can reproduce this bug.

import mxnet as mx
from mxnet.test_utils import rand_ndarray

def testBug(dev):
    shape_lhs = (200, 200)
    shape_rhs = (200, 200)
    mx_sparse = rand_ndarray(shape_lhs, 'csr', density=0.01).as_in_context(dev)
    mx_dns = rand_ndarray(shape_rhs, 'default', density=1.0).as_in_context(dev)
    mx.nd.dot(mx_sparse, mx_dns, transpose_a=True, transpose_b=False
              , forward_stype='default')
    mx.nd.waitall()
    
if __name__ == "__main__":
    print('test dot(csr.T, dns)=dns on cpu')
    testBug(mx.cpu())
    print('test dot(csr.T, dns)=dns on gpu')
    testBug(mx.gpu())

Here is the log info. It tells the storage types of dot(csr.T, dns)=dns would fallback, and then the actually running code is dot(dns,dns)=dns

test dot(csr.T, dns)=dns on cpu
[21:56:09] src/operator/nn/./../../common/utils.h:416: 
Storage type fallback detected:
operator = dot
input storage types = [csr, default, ]
output storage types = [default, ]
params = {"forward_stype" : default, "transpose_b" : False, "transpose_a" : True, }
context.dev_mask = cpu
The operator with default storage type will be dispatched for execution. You're seeing this warning message because the operator above is unable to process the given ndarrays with specified storage types, context and parameter. Temporary dense ndarrays are generated in order to execute the operator. You can set environment variable MXNET_STORAGE_FALLBACK_LOG_VERBOSE to 0 to suppress this warning.
test dot(csr.T, dns)=dns on gpu
[21:57:01] src/operator/nn/./../../common/utils.h:416: 
Storage type fallback detected:
operator = dot
input storage types = [csr, default, ]
output storage types = [default, ]
params = {"forward_stype" : default, "transpose_b" : False, "transpose_a" : True, }
context.dev_mask = gpu
The operator with default storage type will be dispatched for execution. You're seeing this warning message because the operator above is unable to process the given ndarrays with specified storage types, context and parameter. Temporary dense ndarrays are generated in order to execute the operator. You can set environment variable MXNET_STORAGE_FALLBACK_LOG_VERBOSE to 0 to suppress this warning.

speed up after fixing the bug

test dot(csr.T, dns)=dns on cpu
1.00 % with fallback: 4.566107, without fallback: 4.191508
0.50 % with fallback: 8.580608, without fallback: 8.105466
0.10 % with fallback: 8.982658, without fallback: 8.352470
0.05 % with fallback: 9.716504, without fallback: 9.003553
0.01 % with fallback: 11.551130, without fallback: 10.684023
test dot(csr.T, dns)=dns on gpu
1.00 % with fallback: 0.574556, without fallback: 0.559032
0.50 % with fallback: 1.073330, without fallback: 1.044696
0.10 % with fallback: 3.619832, without fallback: 3.529394
0.05 % with fallback: 4.883098, without fallback: 4.761823
0.01 % with fallback: 9.730866, without fallback: 9.499284

The benchmark script is here https://github.com/XiaotaoChen/incubator-mxnet/blob/Mytest/example/sparse/temp_test/testBug.py
@pengzhao-intel @TaoLv

[pengzhao-intel]

@haojin2