fix(bicep): stop corrupting hover/definition/references/symbols/formatting/codeAction with completion injections

_ls_to_client called _inject_completions() on EVERY response with an id+result,
not just textDocument/completion responses. pop_context() defaulted to
{'type': 'unknown'} for any untracked id, which _inject_completions treats as
'inject module-name completions' — so list-shaped LS responses for
definition/references/documentSymbol/formatting/codeAction were silently
replaced or prefixed with the Bicep module catalog list instead of returning
real Bicep LangServer results.

Found via a full LSP-capability audit against production: definition,
references, documentSymbol, formatting and codeAction all incorrectly
returned module-completion items instead of real results.

Fix: pop_context() now returns None when the response id wasn't recorded as
a pending completion request, and _ls_to_client only calls
_inject_completions() when a real completion context was found — every other
response now passes through untouched.

tests/test_proxy.py

@@ -502,6 +502,7 @@ def test_session_records_and_pops_context():
     assert ctx["type"] == "param"
     assert ctx["module"] == "appservice"
 
-    # Second pop returns unknown
-    assert session.pop_context(42)["type"] == "unknown"
+    # Second pop (or any untracked id) returns None — the response must pass
+    # through untouched rather than being mistaken for a completion response.
+    assert session.pop_context(42) is None