Faster tensor name formatting

[ikawrakow]

When building the compute graph, tensors get assigned names via ggml_format_name. For small models with very fast inference, a non-negligible fraction of the overall compute time (not just graph building time) is spent in that function. Given the names being assigned, we don't need to use snprintf as ggml_format_name does. This PR replaces calls to ggml_format_name with calls to a new function, which is very simple but still quite a bit faster.

For e.g. Ling-mini-2.0 running on RTX-4080 we get a ~1% TG performance gain.

Here some sweep-bench results with this PR for Ling-mini-2.0

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.142	14420.91	1.220	419.50
2048	512	2048	0.117	17508.91	1.283	398.98
2048	512	4096	0.124	16528.93	1.362	375.81
2048	512	6144	0.129	15815.04	1.447	353.95
2048	512	8192	0.136	15058.71	1.505	340.12
2048	512	10240	0.143	14353.90	1.593	321.36
2048	512	12288	0.149	13723.51	1.644	311.36
2048	512	14336	0.156	13142.36	1.725	296.88
2048	512	16384	0.163	12563.18	1.780	287.61
2048	512	18432	0.170	12012.86	1.845	277.50
2048	512	20480	0.176	11605.83	1.929	265.47
2048	512	22528	0.184	11130.56	1.987	257.67
2048	512	24576	0.189	10838.27	2.074	246.81
2048	512	26624	0.197	10376.34	2.129	240.44
2048	512	28672	0.204	10015.94	2.193	233.44
2048	512	30720	0.209	9791.92	2.273	225.25

This is with grouped expert routing enabled (-ger). Compared to #838, where grouped expert routing was added on CUDA, at low context TG performance has improved by ~11% and PP performance by 8%. This is the cumulative effect of this PR, and PRs #858, #853, #845, #840.