Return self in InferInput setters for chain-initialization

[GuanLuo]

UX improvement.

[rmccorm4]

Nice addition.

I've been thinking changing our constructor from something like this:

class InferInput:
    def __init__(self, name, shape, datatype):

to something along the lines of this:

class InferInput:
    def __init__(self, name, shape=None, datatype=None, tensor=None):
      if shape:
        self.shape = shape
      if datatype:
        self.datatype = datatype
        
      # Improve existing numpy support to just take an array to initialize input:
      if isinstance(tensor, np.ndarray):
        if not shape:
          self.shape = tensor.shape
        if not datatype:
          self.datatype = utils.np_to_triton_dtype(tensor.dtype)
        self.set_data_from_numpy(tensor)
        
      # Future torch tensor support - might be better to have an explicit InferInput.FromTorch()
      # function here that does `import torch` dynamically so we don't require torch
      if isinstance(tensor, torch.tensor):
        ...

or some wrapper since the above "Either/Or" requirement of shape+datatype or tensor isn't very clear:

class InferInputWrapper:
    def __init__(self, name, tensor):
      # Improve existing numpy support to just take an array to initialize input:
      if isinstance(tensor, np.ndarray):
        if not shape:
          self.shape = tensor.shape
        if not datatype:
          self.datatype = triton_to_np(tensor.dtype)
        self.set_data_from_numpy(tensor)
        
      ...

Then usage for non-shared memory case can simply become something like this:

data = np.zeros((1, 16), dtype=np.float32)
inputs = [InferInput("INPUT0", data)]