MNT: Move constructors to a direct style

[mtsokol]

Hi @hameerabbasi,

This PR moves MLIR constructors to the direct style (with: sparse-assembler{direct-out=true}).
I think it's more concise with this approach - _constructors.py LOC reduced by 50%, to 250.

Now all Tensors created from NumPy arrays or SciPy sparse arrays are owned and freed by the python side.

The ones that are results of MLIR ops (like add right now) are allocated by MLIR and owns_memory in Tensor class is set to False.

I haven't figured out yet how to correctly free MLIR allocated memrefs, (tried with libc.free) as I think there's still a memory leak after sparse.add(tensor_a, tensor_b).

[hameerabbasi]

Thanks @mtsokol! LGTM so far!

1. Let's make all members of Tensor private.
2. Let's store a list of free-able memref strs. So, for example, in to_sps let's store ["data"] for dense, then we will need to call libc.free on data.allocated. This will apply when we take ownership from MLIR only.
3. We need to store a reference to NumPy/SciPy objects when we create Tensors from them, as we were doing before. The pattern for _hold_self_ref_in_ret will apply still in this case, except we can just use Tensor._base which can be the SciPy or NumPy object, or None.

[codspeed-hq]

CodSpeed Performance Report

Merging #773 will degrade performances by 51.96%

_{Comparing direct-style (ef3e62e) with main (47b246b)}

Summary

❌ 2 (👁 2) regressions
✅ 338 untouched benchmarks

Benchmarks breakdown

	Benchmark	main	direct-style	Change
👁	test_index_fancy[side=100-rank=1-format='coo']	680 µs	1,415.4 µs	-51.96%
👁	test_index_slice[side=100-rank=2-format='gcxs']	2.2 ms	2.5 ms	-12.61%

[mtsokol]

@hameerabbasi I applied all your review comments.

[hameerabbasi]

A simplification to the memory logic: We can do most of this in __del__, check for self._owns_memory and then call libc.free on the constituent memrefs if so.

[hameerabbasi]

Thanks for the 🥚-cellent work, @mtsokol! LGTM.