Falcon7b prefill remaining fixes and cleanup after enabling optimized version

[s-jovic]

• Segfault for 1k and 2k prefill on multi-device (single run) and single device (when run in a loop): Seg fault in falcon7b prefil optimised attention #8644
• Async mode (not all ops used in optimized prefill support async mode)
• Make model initialization agnostic to sequence length
• GS path - decide whether to use the optimized version for GS and if so make it work
• Use appropriate memory config configuration (l1 sharded?)
• Resolve lm head e2e perf impact
• Push PCC to 0.99: [issue](Falcon7b prefill PCC below 0.99 #8487)
• Reduce memory usage of e2e tests - the 2k test uses up to 80gb ram for some reason; try to run PyTorch and tt model sequentially

[pavlepopovic]

• Check that all MMs have optimal settings (sunblock_h/w,, in0_block_w) following di/dt fixes (attention, MLP, LM head matmuls)
• Check if any more sharding is possible throughout model

[s-jovic]

• remove persistent kernel cache usage
• put optimized attention on CI
• unify paths for 128/1024/2k and other sequence lengths
• perf breakdown
• update perf targets in CI tests
• add multi-chip 128, 1k, 2k prefill tests