Update on "[2/x] clean up casting functions: delayed scaling"

Summary:

Removes delayed scaling from `float8_tensor.py`. After this PR, the
invariant is that everything in `float8_tensor.py` requires the scale to
be calculated elsewhere. This moves the codebase towards separation of
concerns for calculating the scale (via various scaling strategies),
separated from creating an instance of `Float8Tensor`.

Note that stateful delayed scaling is the reason we need this separation.

Test Plan:

```
./test/test_everything.sh
```

Reviewers:

Subscribers:

Tasks:

Tags:

Differential Revision: [D60291447](https://our.internmc.facebook.com/intern/diff/D60291447)

[ghstack-poisoned]

test/test_compile.py

@@ -20,9 +20,9 @@
     get_float8_layers,
     sync_float8_amax_and_scale_history,
 )
+from float8_experimental.float8_scaling_utils import cast_to_float8_delayed
 from float8_experimental.float8_tensor import LinearMMConfig
 from float8_experimental.float8_utils import e4m3_dtype
-from float8_experimental.float8_scaling_utils import cast_to_float8_delayed
 
 from torch._dynamo.test_case import TestCase as DynamoTestCase
 from torch._dynamo.testing import CompileCounterWithBackend