Fix Phi3 128k finally: use position ids to switch between short/long scaling

[EricLBuehler]

Before, we used the start offset. This is not correct. Now, we use position ids based on the context length and past KV length.

[github-actions]

Code Metrics Report

  ───────────────────────────────────────────────────────────────────────────────
Language                 Files     Lines   Blanks  Comments     Code Complexity
───────────────────────────────────────────────────────────────────────────────
Rust                        70     23594     1555       510    21529       1309
───────────────────────────────────────────────────────────────────────────────
Total                       70     23594     1555       510    21529       1309
───────────────────────────────────────────────────────────────────────────────
Estimated Cost to Develop 78,063
Estimated Schedule Effort 11.865794 months
Estimated People Required 5.076793
───────────────────────────────────────────────────────────────────────────────
Processed 780091 bytes, 0.780 megabytes (SI)
───────────────────────────────────────────────────────────────────────────────