[jit] Fix future type annotation in python

[wanchaol]

We should fix the Future type annotation to make it also work in python, fork will immediately execute the function single threaded and returns a Future of the result, which wait just unpacks.(the same way that tracing did).

from torch import nn
from typing import List
import torch

class Encoder(nn.Module):
    def forward(self, x):
        return x

class EncoderEnsemble2(nn.Module):
    def __init__(self, encoders : List[nn.Module]):
        super().__init__()
        self.encoders = nn.ModuleList(encoders)

    def forward(self, x):
        futures = torch.jit.annotate(
            List[Future[Tensor]], []
        )
        for encoder in self.encoders:
            futures.append(
                torch.jit._fork(encoder, x)
            )

        all_outputs = []
        for future in futures:
            all_outputs.append(torch.jit._wait(future))
        return all_outputs

model =EncoderEnsemble2([Encoder()])
output = model(torch.randn(3, 4))

cc @ezyang @gchanan @zou3519 @suo