Fix errors in ttnn.nlp_create_qkv_heads_decode when input shard is bigger than one tile

[kpaigwar]

Problem description

Presently, CreateQKV Decode Op is only tested with hardcoded input shard shape (32, 32).
When relaxing this constraint and moving to bigger shard (32, 128). This results into two issues

1. PCC issues when input is on contiguous core_grid
2. FATAL error in creating output_specs when input is on subcoregrids.

What's changed

1. Fixed bug in reader/writer kernel when calculating num_q_cores and num_kv_cores.
2. Simplified create_output_specs, now fetching available output core_grid from output memory config.
3. Automatic calculation of output core_grid, shard_shape, BufferType when no output_memory_config is specified.

Checklist

• All post commit CI passes

[cglagovichTT]

please update PR title and description with more detail