Fix /ctx error handling in interactive REPL

[superbling]

Summary

Fix two issues with the /ctx command that caused the interactive REPL to terminate unexpectedly.

1. /ctx with an invalid value kills the process. The REPL handler called parse_int(), which calls exit(2) on bad input. That is correct for command-line argument parsing, but in the REPL it drops the KV cache and the entire conversation on a typo (e.g. /ctx abc). Added a parse_int_repl() twin that prints the same error and returns false, so the REPL can simply prompt again.

2. /ctx with a valid value but failing session creation also kills the chat. repl_chat_set_ctx() freed the existing session before trying to create the replacement. If creation failed (e.g. allocation error for the new context size), the old session was already gone and run_repl broke out of the loop. Dropped the redundant pre-free; repl_chat_create_session() already keeps the old session intact on failure, so /ctx now reports the error and keeps the current context.

Test plan

• make — clean build, no new warnings
• Piped commands into ./ds4 -c 4096:
  • /ctx abc → prints ds4: invalid value for /ctx: abc, REPL continues (previously: process exit)
  • /ctx 4096 → successfully recreates the session (logs new context buffers)
  • /quit → exits cleanly

[superbling]

Closing in favor of two smaller PRs that split this work apart:

• Don't kill the REPL on /ctx with invalid input #54 — Don't kill the REPL on /ctx with invalid input (pure bug fix, no behavioral change)
• Keep the chat session when /ctx fails to recreate it #55 — Keep the chat session when /ctx fails to recreate it (includes a behavioral change worth its own review)

This original PR bundled both, but the second piece is really a design call about whether /ctx failure should terminate the REPL or stay at the prompt — different in nature from the parse-error fix. Splitting them lets the unambiguous fix land independently.

Sorry for the noise.