Remove deprecated lazy tensor

[saitcakmak]

The lazy tensor functionality has been deprecated (and split off as linear operator) for over two years. Since then, the old lazy tensor classes have been raising a deprecation warning and pointing to the corresponding linear operator classes via the deprecated_lazy_tensor helper class.

While profiling candidate generation in Ax / BoTorch, I've noticed 5-15% of the runtime being spent in deprecated_lazy_tensor.__init__. This class is being instantiated hundreds of thousands of times during each candidate generation, adding a significant overhead without a clear benefit. I think it's time to remove it and claw back that overhead.

Here's a sample of an example cProfile output:

         68404451 function calls (65915300 primitive calls) in 150.665 seconds
   Ordered by: internal time
   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
     1876   42.210    0.023   42.210    0.023 {method 'run_backward' of 'torch._C._EngineBase' objects}
   725407    8.305    0.000   15.168    0.000 functional.py:79(broadcast_shapes)
19530046/18762961    4.030    0.000    5.435    0.000 {built-in method builtins.isinstance}
    46998    3.955    0.000    3.955    0.000 {method 'matmul' of 'torch._C.TensorBase' objects}
  3914524    2.738    0.000    3.857    0.000 symbolic_shapes.py:393(guard_size_oblivious)
   338423    2.047    0.000   14.049    0.000 module.py:1932(__setattr__)
   368383    1.844    0.000    9.709    0.000 _linear_operator.py:155(__init__)
    90242    1.811    0.000    4.984    0.000 inspect.py:2268(_signature_from_function)
    73322    1.732    0.000    1.732    0.000 {built-in method torch.cat}
    22560    1.610    0.000    4.713    0.000 exact_prediction_strategies.py:365(exact_predictive_covar)
    20658    1.548    0.000    4.651    0.000 safe_math.py:422(fatmoid)
**693531/298896    1.546    0.000   22.865    0.000 lazy_tensor.py:32(__init__)**

You can see that deprecated_lazy_tensor.__init__ was called 693531 times, resulting in a 22.865 second overhead in a total runtime of 150.665 seconds.

[Balandat]

lgtm - cc @gpleiss, @jacobrgardner, @JonathanWenger