Add stream context

[cyx-6]

This PR adds the stream context into ffi, so that ffi env stream can be updated. The tvm_ffi.use_torch_stream is for wrapping the torch stream/graph context. And lower-level tvm_ffi.use_raw_stream is for creating context with existing stream handle.

Example for tvm_ffi.use_torch_stream:

case with torch stream:

stream = torch.cuda.Stream()
stream_context = torch.cuda.stream(stream)
with tvm_ffi.use_torch_stream(stream_context):
  ...

case with torch cuda graph

graph = torch.cuda.CUDAGraph()
graph_context = torch.cuda.graph(graph)
with tvm_ffi.use_torch_stream(graph_context):
  ...

case with current stream by default

stream = torch.cuda.Stream()
stream_context = torch.cuda.stream(stream)
with torch.cuda.stream(stream):
  with tvm_ffi.use_torch_stream():
    ...

Eaxmple for tvm_ffi.use_raw_stream:

device = tvm_ffi.device(...)
stream_handle = ...
with tvm_ffi.use_raw_stream(device, stream_handle):
  ...

[tqchen]

need indent one level and a blank line between code block and code

[tqchen]

given this is not cython dependent, move out to stream.py

[tqchen]

we can expose this as def, so it can be called from python side

[tqchen]

new line not needed?