Fix einsum dropping a trailing empty subscript

[Adityaj0]

Fixes #4298.

Proposed changes

An empty subscript denotes a scalar operand, but parse() tokenized the left hand side with getline:

std::stringstream ss(lhs);
while (getline(ss, token, ',')) {
  input_list.push_back(token);
}

getline stops at the last delimiter, so a trailing empty field is never produced. "i," parsed as ["i"] instead of ["i", ""], and the operand-count check then rejected the call:

mx.einsum("i,->i", mx.zeros((3,)), mx.array(2.0))
# ValueError: [einsum] Number of operands, 2, does not match number of input subscripts, 1

A leading empty subscript worked, because getline does emit an empty first field, which is why ",i->i" was accepted while "i,->i" was not. That asymmetry is what makes it look like a tokenizer bug rather than an intentional restriction.

This splits on commas directly so every field is kept, including trailing and all-empty ones.

equation	NumPy	before	after
",i->i"	(3,)	(3,)	(3,)
"i,->i"	(3,)	error	(3,)
",->"	()	error	()
"->"	()	error	()
"ij,->ij"	(2,3)	error	(2,3)
",,->"	()	error	()

The validation that rejects genuinely bad input is untouched: the operand-count check and the per-operand in.size() != operands[i].ndim() check still run, so the existing negative cases in tests/einsum_tests.cpp continue to throw:

CHECK_THROWS(einsum("", {}));                 // 1 subscript, 0 operands
CHECK_THROWS(einsum("", {array({1, 2})}));    // "" requires a 0-d operand

Tests

Added test_scalar_operands to python/tests/test_einsum.py covering scalar operands in leading, trailing and middle positions, all-scalar equations, and "->", each checked against NumPy. It also asserts the negative cases still raise: mismatched operand count, and an empty subscript paired with a non 0-d operand.

• I have read the CONTRIBUTING document
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (not needed, no API change)
• I have run pre-commit run --all-files to format my code and installed pre-commit prior to committing changes

Verified with a CPU-only build (-DMLX_BUILD_METAL=OFF):

• C++ suite: 247 test cases, 3326 assertions, all pass
• Python: test_einsum, test_ops, test_autograd, test_vmap, test_linalg, test_blas, test_nn, test_array, test_reduce (464 tests) pass

This is independent of #4125, which touches batch_tensordot in the same file but a different function.