Validate input rank in the tf2xla QR and SVD kernels

[vishwakt]

Fixes #110798.

Problem

The tf2xla kernels for Qr, Svd, and XlaSvd pass their input straight into the XLA builder libraries without validating its rank. The graph-level shape functions reject known-rank inputs below rank 2, but an input whose static rank is unknown, such as the reporter's StackPopV2 output or any unknown-rank placeholder, reaches the kernels unchecked at XLA compile time:

• QROp::Compile calls xla::QrExplicit, which reads dimension rank - 2 = -1 and dies on the fatal ShapeUtil::GetDimensionNumber check (shape_util.cc:952, the check in the report).
• SvdOp::Compile dies one step earlier, on its own input_shape.dim_size(dims() - 2) call, hitting the d >= 0 check in TensorShape::dim_size. This sibling was found while verifying the report and is fixed here as well; XlaSvdOp shares the same gap.

Both reproduce on XLA CPU with jit_compile=True (verified locally, exit 134 with the reported check-failure text for QR and the TensorShape variant for SVD). The neighboring Cholesky and SelfAdjointEig lowerings already reject rank-1 inputs with clean InvalidArgumentErrors, which confirms the intended behavior for this input class.

Fix

Validate rank >= 2 in all three kernels before any trailing dimensions are read, using the validation and message style of the neighboring MatrixSolve kernel ("Input must have rank >= 2, got shape ..."). Compilation now fails with a regular InvalidArgumentError instead of aborting the process.

Tests

testVectorInputRaisesError in compiler/tests/qr_op_test.py and compiler/tests/svd_op_test.py feeds a rank-1 value through an unknown-rank placeholder, which is exactly the path that bypasses graph-level shape inference; on unfixed builds this aborts the test process, and with the fix it asserts the InvalidArgumentError. BUILD dependencies updated for the strict-deps targets (kernel targets gain the absl status and strings deps used by the new checks, test targets gain //tensorflow/python/framework:errors).