Add support to non-fp32 input dtype in level3/33_VanillaRNN.py and level3/35_LSTM.py

[mzweilin]

This PR fixes a bug in two problems level3/33_VanillaRNN.py and level3/35_LSTM.py that Model uses hard-coded FP32 tensors.

It is related to #79 and #80. You will find the bug if you try to run the agent on the two problems.

Please consider updating the dataset in HuggingFace too: https://huggingface.co/datasets/ScalingIntelligence/KernelBench

An alternative fix could be moving the random tensors from Model to get_inputs().

How to reproduce the bug

$ python -i KernelBench/level3/33_VanillaRNN.py

init_inputs = get_init_inputs()
inputs = get_inputs()
m = Model(*init_inputs)
output = m(*inputs)
m_bf16 = m.to(dtype=torch.bfloat16)
inputs_bf16 = [inp.to(dtype=torch.bfloat16) for inp in inputs]
output_bf16 = m_bf16(*inputs_bf16)

RuntimeError: mat1 and mat2 must have the same dtype, but got Float and BFloat16

$ python -i KernelBench/level3/35_LSTM.py

init_inputs = get_init_inputs()
inputs = get_inputs()
m = Model(*init_inputs)
output = m(*inputs)
m_bf16 = m.to(dtype=torch.bfloat16)
inputs_bf16 = [inp.to(dtype=torch.bfloat16) for inp in inputs]
output_bf16 = m_bf16(*inputs_bf16)

RuntimeError: could not create a primitive descriptor for the LSTM forward propagation primitive. Run workload with environment variable ONEDNN_VERBOSE=all to get additional diagnostic information.