cartesian: Extend 2D temporaries to `gt:X` backend

[FlorianDeconinck]

Following this PR we have introduced 2D temporaries as described in the ADR.

gt:X backends were excluded from this PR because GridTools pre-suppose that all computation of temporaries are done on the 3D grid at least - and won't allocate below that dimensionality. This is a fair limitation since GridTools doesn't provide the guardrails against race condition that GT4Py does.

We can circumvent this by creating a pool of buffers passed as arguments. In details a stencils written like this

@stencil(backend=..., dtype={"My2DType": Field[IJ, np.float64]})
def the_stencil(A: Field[IJK, np.float64], B: Field[IJK, np.float64]):
    with computation(FORWARD), inteval(...):
        tmp_2D: Field[IJ, np.float64] = 0
        ...

would become during OIR -> GTCPP

@stencil(backend=..., dtype={"My2DType": Field[IJ, np.float64]})
def the_stencil(A: Field[IJK, np.float64], B: Field[IJK, np.float64], tmp_2D: Field[IJ, np.float64]):
    with computation(FORWARD), inteval(...):
        tmp_2D = 0
        ...

And the tmp_2D would be sourced from a pooled memory of 2D temporaries properly scoped to the IJ required - with all information generated and stored on the stencil compiled during codegen of GTCPP

Code would inject it right before the pybind11 bindings.

A pass would need to be done to make sure the temporaries don't re-use the name.

[FlorianDeconinck]

Another option - more GridTools centric - would be to turn any 2D temporaries to 3D using a double "fill pass" e.g. every write to a field would look like this

with computation(FORWARD):
    with interval(0, 1):
        tmp_2d = bb
    with interval(1, -1):
        tmp_2d = tmp_2d[0, 0, -1]

This would allow GridTools K-Caching to go in play

We could generate this automatically.