fix: refuse per-axis int8 activations; convert outermost-axis concatenation

[asteinh]

• The compiler refuses an int8 activation quantized per axis, for every operator. A plan record holds one scale and zero point for an activation, and readers now ignore the producer's per-channel requantization it may carry, so a per-axis encoding could not be told apart. Before, some runtime readers rejected it and others read only its first scale.
• The TFLite frontend converts CONCATENATION and PACK on the outermost axis and at ranks other than 3 and 4, by joining the parts as contiguous runs. Keras unrolls an LSTM with PACK on axis 0.
• New int8 and float32 fixtures, bit-exact against TFLite Micro: convolution, depthwise and fully connected with non-zero biases; a per-channel convolution feeding a split; outermost-axis concatenation and pack; rank-2 concatenation; and an unrolled Keras LSTM.

Needs the runtime branch of the same name.