Add a16w8, the QNN 16-bit-activation / 8-bit-weight precision

Say why, not what went wrong

Point the compatibility link at the renamed docs page

Add the unshipped-backends generated section

Add the model card standard

Stop storing quantized and default on a variant

Drop the dead MODEL_SPEC.md reference

Let a variant's methods nest under its components

Update config.schema.json

Relax tensorSpec.name to optional

Add text-to-speech, text-detection, text-recognition capabilities

Make `methods` optional (.pte introspection is a follow-up step)

Initial spec: config.schema.json, precisions.json, README

initial commit