Add more manual implementations for TF's einsum

[james77777778]

We need these new subscripts to utilize int8,int8->int32 einsum for backprop in #19356

                inputs_grad = ops.einsum(
                    # self._custom_gradient_equation is '{output_spec},{weight_spec}->{input_spec}'
                    self._custom_gradient_equation, upstream, float_kernel
                )

To find out all of these subscripts:

import re

from keras.backend.tensorflow.numpy import _normalize_einsum_subscripts

currently_supported_subscripts = [
    "a,b->ab",
    "ab,bc->ac",
    "abc,cd->abd",
    "abcd,abed->abce",
    "abcd,adbe->acbe",
    "abcd,aecd->acbe",
    "abcd,aecd->aceb",
    "abc,cde->abde",
    "abc,dce->abde",
    "abcd,cde->abe",
    "abcde,aebf->adbcf",
    "abcde,afce->acdbf",
]


def custom_gradient_subscripts(equation):
    dot_replaced_string = re.sub(r"\.\.\.", "0", equation)
    split_string = re.match(
        "([a-zA-Z]+),([a-zA-Z]+)->([a-zA-Z]+)", dot_replaced_string
    )
    if split_string is not None:
        input_spec = split_string.group(1)
        weight_spec = split_string.group(2)
        output_spec = split_string.group(3)
    return f"{output_spec},{weight_spec}->{input_spec}"


possible_subscripts = []
for s in currently_supported_subscripts:
    t = _normalize_einsum_subscripts(custom_gradient_subscripts(s))
    possible_subscripts.append(t)

supported_subscripts = set(currently_supported_subscripts)
possible_subscripts = set(possible_subscripts)
missing_subscripts = possible_subscripts - supported_subscripts
print(sorted(missing_subscripts))

# ['ab,b->a', 'ab,cb->ac', 'abc,dc->abd', 'abc,dec->abde', 'abcd,abde->abce',
# 'abcd,acbe->adbe', 'abcd,ced->abe', 'abcd,ecd->abe']

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 90.32258% with 6 lines in your changes are missing coverage. Please review.

Project coverage is 75.86%. Comparing base (aa3a61b) to head (edcaa97).
Report is 2 commits behind head on master.

Files	Patch %	Lines
keras/backend/tensorflow/numpy.py	90.32%	3 Missing and 3 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #19368   +/-   ##
=======================================
  Coverage   75.86%   75.86%           
=======================================
  Files         366      366           
  Lines       40417    40479   +62     
  Branches     7855     7869   +14     
=======================================
+ Hits        30661    30711   +50     
- Misses       8061     8068    +7     
- Partials     1695     1700    +5     

Flag	Coverage Δ
keras	75.72% <90.32%> (+<0.01%)	⬆️
keras-jax	60.02% <0.00%> (-0.10%)	⬇️
keras-numpy	54.29% <0.00%> (-0.09%)	⬇️
keras-tensorflow	61.32% <90.32%> (+0.04%)	⬆️
keras-torch	60.31% <0.00%> (-0.11%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[fchollet]

LGTM, thank you.