appendToken
Append a single token's K/V projections for one layer.
The runtime calls this once per layer per generated token. The cache is responsible for encoding/compressing the data according to keyEncoding and valueEncoding.
Parameters
layer
Layer index (0-based)
key
Key projection numHeads, headDim as float
value
Value projection numHeads, headDim as float
Throws
if the cache is full (currentSeqLen>= maxSeqLen)