[CI] Pin h5py version to < 3.0 to workaround issues with TF/Keras

[leandron]

h5py recently released version 3.0 (dates can be seen here: https://pypi.org/project/h5py/#history). This breaks TF 2.3 and Keras, as reported already by others, e.g. tensorflow/tensorflow#44467.

Because of that, there are a few tests that fail:

FAILED tests/python/frontend/keras/test_forward.py::TestKeras::test_forward_resnet50[keras]
FAILED tests/python/frontend/keras/test_forward.py::TestKeras::test_forward_mobilenet[keras]
FAILED tests/python/frontend/keras/test_forward.py::TestKeras::test_forward_resnet50[tf_keras]
FAILED tests/python/frontend/keras/test_forward.py::TestKeras::test_forward_mobilenet[tf_keras]
ERROR tests/python/driver/tvmc/test_compiler.py::test_compile_keras__save_module
ERROR tests/python/driver/tvmc/test_compiler.py::test_cross_compile_aarch64_keras_module
ERROR tests/python/driver/tvmc/test_frontends.py::test_load_model__keras - At...
ERROR tests/python/driver/tvmc/test_frontends.py::test_load_model___wrong_language__to_tflite

The reason of the failures/errors indeed is the same as reported upstream:

    def load_weights_from_hdf5_group(f, layers):
      """Implements topological (order-based) weight loading.
    
      Arguments:
          f: A pointer to a HDF5 group.
          layers: a list of target layers.
    
      Raises:
          ValueError: in case of mismatch between provided layers
              and weights file.
      """
      if 'keras_version' in f.attrs:
>       original_keras_version = f.attrs['keras_version'].decode('utf8')
E       AttributeError: 'str' object has no attribute 'decode'

/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/saving/hdf5_format.py:660: AttributeError

To workaround that, this PR:

• Pin h5py to use the previous major release (2.x) and not new version 3.0, due to incompatibilities with TF and Keras that make TVMC and Frontend tests to fail

Note: just to make it clear, this is caused by the h5py release and affects TF >= 2.1.0 so it is not related in any way with the recent update to TF 2.3.1.

cc @tqchen @u99127 @areusch @ANSHUMAN87

[leandron]

Thanks @tqchen for merging this today 👍