[Torchscript] Adds Sequence and Text feature postprocessing

[geoffreyangus]

This PR implements Sequence and Text postprocessing. In order to do this, we additionally standardize the outputs of both Sequence and Text output features. Now, both features output the following:

• prediction: the maximum token for each timestep
• probabilities: the probability of the maximum token for each timestep
• probability: the log-probability of the predicted sequence

{
    "predictions": [
        ["lQhz", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB"],
        ["lQhz", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB", "HxgtdB"],
    ],
    "probabilities": tensor(
        [
            [0.2930, 0.2792, 0.2555, 0.2349, 0.2203, 0.2100, 0.2030, 0.1983, 0.1954],
            [0.2899, 0.2953, 0.2627, 0.2376, 0.2212, 0.2103, 0.2031, 0.1984, 0.1955],
        ]
    ),
    "probability": tensor(
        [
            -13.2349,
            -13.1438,
        ]
    ),
}

[github-actions]

Unit Test Results

       6 files  ±0         6 suites  ±0   2h 13m 29s ⏱️ + 9m 41s
2 831 tests +1  2 797 ✔️ +2    34 💤 ±0  0 ❌  - 1 
8 493 runs  +3  8 387 ✔️ +4  106 💤 ±0  0 ❌  - 1 

Results for commit 4c450d2. ± Comparison against base commit f43d23e.

♻️ This comment has been updated with latest results.

[justinxzhao]

Nice consolidation. Had a few minor comments.