[Performance] Random speedups

[albanD]

This is a proposal for speedup improvements for TensorDict struct by adding fast path for common use cases.
I only tested test/test_tensordict.py locally (the subset that works on main), there is most likely a few things I missed!

Runtime is tested with

from torch.nn import Transformer
from tensordict import TensorDict
import torch.utils.benchmark
t = Transformer()
td = TensorDict.from_module(t)

print(torch.utils.benchmark.Timer("td.flatten_keys()", globals=globals()).adaptive_autorange())
td.lock_()
print(torch.utils.benchmark.Timer("td.flatten_keys()", globals=globals()).adaptive_autorange())
print(torch.utils.benchmark.Timer("""td.flatten_keys(".")""", globals=globals()).adaptive_autorange())

Result before:

<torch.utils.benchmark.utils.common.Measurement object at 0x7f5c966e6a90>
td.flatten_keys()
  Median: 477.80 us
  IQR:    2.82 us (476.36 to 479.17)
  4 measurements, 1000 runs per measurement, 1 thread
<torch.utils.benchmark.utils.common.Measurement object at 0x7f5c95c67910>
td.flatten_keys()
  Median: 524.48 ns
  IQR:    3.03 ns (523.41 to 526.44)
  4 measurements, 100000 runs per measurement, 1 thread
<torch.utils.benchmark.utils.common.Measurement object at 0x7f5c95c677d0>
td.flatten_keys(".")
  Median: 727.95 ns
  IQR:    8.25 ns (724.57 to 732.82)
  4 measurements, 100000 runs per measurement, 1 thread

Result after:

<torch.utils.benchmark.utils.common.Measurement object at 0x7f1811c61250>
td.flatten_keys()
  Median: 226.33 us
  IQR:    0.10 us (226.25 to 226.34)
  4 measurements, 1000 runs per measurement, 1 thread
<torch.utils.benchmark.utils.common.Measurement object at 0x7f1810d1b1d0>
td.flatten_keys()
  Median: 159.65 ns
  IQR:    0.13 ns (159.57 to 159.70)
  4 measurements, 1000000 runs per measurement, 1 thread
<torch.utils.benchmark.utils.common.Measurement object at 0x7f1810d1aa10>
td.flatten_keys(".")
  Median: 216.46 ns
  IQR:    0.56 ns (216.21 to 216.77)
  4 measurements, 1000000 runs per measurement, 1 thread

[vmoens]

For some reason the benchmarks don't appear in this PR but you can check them here
they look pretty good! Thanks so much!