fix(#69): fix tensorInitWithDistribution buffer overflow and add rank…

[LeoBuron]

… check to doDimensionsMatch

tensorInitWithDistribution computed the element count by summing dimensions instead of multiplying them (+= instead of *=, init 0 instead of 1). This caused two simultaneous failures:

1. Buffer overflow for bias tensors with shape [1, N]: sum = 1+N > product = N, so memset wrote one float past the end of the caller's data buffer.

2. Under-initialization for weight tensors with shape [M, N] where M,N > 1: sum = M+N << product = M*N, so only a fraction of values received their intended distribution (e.g. Xavier), with the rest staying at zero.

The overflow alone would not explain the crash, but this codebase uses zero-copy shape storage: initTensorWithQFloat calls setShape(shape, dims, ...) which stores shape->dimensions = dims — a raw pointer into the caller's stack frame, not a deep copy. This is a deliberate design choice to minimize heap allocations on MCU targets where memory is scarce.

The consequence is that the 4-byte memset overflow from bias0Data into the adjacent bias0Dims array on main()'s stack corrupted dims[0] from 1 to 0. Because the tensor's shape points directly at that array, the bias permanently read shape [0, N] instead of [1, N]. The forward pass then hit doDimensionsMatch comparing output [1, N] against bias [0, N], producing the "Dim 1: 0, Dims 2: 20" error from the issue.

Additionally, doDimensionsMatch did not verify that both tensors have the same rank (numberOfDimensions). When ranks differed, the loop iterated using the first tensor's rank and read out of bounds from the second tensor's VLA. This is now a hard error (exit) rather than silent undefined behavior.

Also:

• MnistExperiment: extracted model construction into buildModel(), added CSV header row. Data arrays are declared static so they outlive the function — required by the zero-copy shape/data pointer design.

Tests added:

• UnitTestTensorApi: 4 tests verifying tensorInitWithDistribution initializes the correct number of elements (product, not sum) for ZEROS, ONES, NORMAL
• UnitTestArithmetic: 2 tests for doDimensionsMatch (same shape, different dims)
• UnitTestMultiLayerTraining: 3 integration tests exercising a 4-layer model (Linear→ReLU→Linear→Softmax) with CrossEntropy loss — the exact architecture that crashed in MnistExperiment