Fixes pytorch#5611.

THCTensor_(baddbmm) assumes that newContiguous will always return a new tensor (this is a bad assumption). At the end of the function, tensors are freed if tensor_new != tensor_old. As a result, some tensors aren't freed if they were initially contiguous and newContiguous is called on them.

Test Plan
code reading
run the following (from the pytorch#5611 bug report) and assert that the memory doesn't leak anymore
import subprocess
import torch
from torch.autograd import Variable

# This is from https://discuss.pytorch.org/t/access-gpu-memory-usage-in-pytorch/3192/4
def get_gpu_memory_map():
    """Get the current gpu usage.

    Returns
    -------
    usage: dict
        Keys are device ids as integers.
        Values are memory usage as integers in MB.
    """
    result = subprocess.check_output(
        [
            'nvidia-smi', '--query-gpu=memory.used',
            '--format=csv,nounits,noheader'
        ], encoding='utf-8')
    # Convert lines into a dictionary
    gpu_memory = [int(x) for x in result.strip().split('\n')]
    gpu_memory_map = dict(zip(range(len(gpu_memory)), gpu_memory))
    return gpu_memory_map

l, m, n = 1, 9, 1
w = torch.nn.Parameter(torch.Tensor(1024, 2, l, m).cuda())
for i in range(10000):
    a = Variable(torch.Tensor(1024, 2, m, n).cuda())
    torch.matmul(w, a).permute(0, 3, 1, 2).mean().backward()
    if i % 100 == 0:
        gpu_mem = get_gpu_memory_map()
        print("GPU: {:.2f} KB".format(gpu_mem[0]))