Preparation for removing by-default import of non-essential packages.

This removes an indirection through rmm.rmm, to simplify removal of
the top-level imports by default.

This will make importing RMM faster if you don't want the hooks for
cupy, numba, or pytorch.

Before, a sampling trace of `import rmm` with pyinstrument shows:

    $ pyinstrument -i 0.01 importrmm.py

      _     ._   __/__   _ _  _  _ _/_   Recorded: 10:19:56  Samples:  67
     /_//_/// /_\ / //_// / //_'/ //     Duration: 0.839     CPU time: 0.837
    /   _/                      v4.4.0

    Program: importrmm.py

    0.839 <module>  importrmm.py:1
    └─ 0.839 <module>  rmm/__init__.py:1
       ├─ 0.315 <module>  rmm/allocators/torch.py:1
       │  └─ 0.315 <module>  torch/__init__.py:1
       │        [96 frames hidden]  torch, <built-in>, enum, inspect, tok...
       ├─ 0.297 <module>  rmm/mr.py:1
       │  └─ 0.297 <module>  rmm/_lib/__init__.py:1
       │     ├─ 0.216 <module>  numba/__init__.py:1
       │     │     [140 frames hidden]  numba, abc, <built-in>, importlib, em...
       │     ├─ 0.040 <module>  numba/cuda/__init__.py:1
       │     │     [34 frames hidden]  numba, asyncio, ssl, <built-in>, re, ...
       │     ├─ 0.030 __new__  enum.py:180
       │     │     [5 frames hidden]  enum, <built-in>
       │     └─ 0.011 [self]  None
       └─ 0.227 <module>  rmm/allocators/cupy.py:1
          └─ 0.227 <module>  cupy/__init__.py:1
                [123 frames hidden]  cupy, pytest, _pytest, attr, <built-i...

After:

    $ pyinstrument -i 0.01 importrmm.py

      _     ._   __/__   _ _  _  _ _/_   Recorded: 10:20:10  Samples:  28
     /_//_/// /_\ / //_// / //_'/ //     Duration: 0.297     CPU time: 0.297
    /   _/                      v4.4.0

    Program: importrmm.py

    0.296 <module>  importrmm.py:1
    └─ 0.296 <module>  rmm/__init__.py:1
       └─ 0.296 <module>  rmm/mr.py:1
          └─ 0.296 <module>  rmm/_lib/__init__.py:1
             ├─ 0.216 <module>  numba/__init__.py:1
             │     [141 frames hidden]  numba, <built-in>, importlib, email, ...
             ├─ 0.040 <module>  numba/cuda/__init__.py:1
             │     [19 frames hidden]  numba, asyncio, ssl, <built-in>, unit...
             ├─ 0.031 [self]  None
             └─ 0.010 __new__  enum.py:180
                   [4 frames hidden]  enum, <built-in>

After this change, a sampling trace of `import rmm` with pyinstrument
shows:

    $ pyinstrument -i 0.01 importrmm.py

      _     ._   __/__   _ _  _  _ _/_   Recorded: 10:37:40  Samples:  9
     /_//_/// /_\ / //_// / //_'/ //     Duration: 0.099     CPU time: 0.099
    /   _/                      v4.4.0

    Program: importrmm.py

    0.099 <module>  importrmm.py:1
    └─ 0.099 <module>  rmm/__init__.py:1
       └─ 0.099 <module>  rmm/mr.py:1
          └─ 0.099 <module>  rmm/_lib/__init__.py:1
             ├─ 0.059 <module>  numpy/__init__.py:1
             │     [31 frames hidden]  numpy, re, sre_compile, <built-in>, s...
             ├─ 0.020 __new__  enum.py:180
             │     [2 frames hidden]  enum
             ├─ 0.010 <module>  ctypes/__init__.py:1
             │     [3 frames hidden]  ctypes, <built-in>
             └─ 0.010 _EnumDict.__setitem__  enum.py:89
                   [3 frames hidden]  enum

Prior to this, when we had expunged cupy and pytorch imports from the
top-level, but not yet numba, an import would take 0.3, rather than
0.1s.