[vMCP] Implement resource exhaustion and DoS protection for session creation

[yrobla]

Depends on #3866

Session-scoped clients multiply backend connections (N sessions × M backends). Without limits, a burst of session creation requests can exhaust file descriptors or memory. Implement configurable limits to protect the system.

Global session limit:

• Configurable via TOOLHIVE_MAX_SESSIONS (or equivalent config key)
• When reached: reject new InitializeRequest with HTTP 503 and a Retry-After header
• Error body: {"error": {"code": -32000, "message": "Maximum concurrent sessions exceeded. Please try again later or contact administrator."}}
• No request queueing — reject immediately; Retry-After gives clients the retry signal
• Do not expose current session count in the error response

Per-client session limit:

• Configurable maximum sessions per client identity/IP
• Enforced independently of the global limit

Idle session timeout:

• Sessions inactive for a configurable period (e.g. 5 minutes with no incoming requests) should be expired proactively, independently of the global session TTL
• This reclaims backend connections faster than waiting for the full TTL to elapse, reducing the steady-state connection count under low-traffic conditions
• Idle timeout must reset on every request touching the session (i.e. any CallTool, ReadResource, or GetPrompt resets the idle clock)
• Configurable separately from the session TTL; idle timeout must be shorter than or equal to TTL

Acceptance Criteria

• A configurable global session limit exists and defaults to a documented value
• Requests exceeding the limit receive HTTP 503 with a Retry-After header
• The error response body does not include the current session count or any internal metrics
• A configurable per-client session limit is enforced independently of the global limit
• A configurable idle timeout exists and defaults to a documented value shorter than the session TTL
• Sessions with no activity for longer than the idle timeout are expired and their backend clients closed
• Any request to a session resets the idle clock
• Idle timeout is enforced independently of — and cannot exceed — the session TTL
• Requests within the limit are unaffected
• Tests cover: 503 + Retry-After on global limit hit, per-client limit enforced, idle session expired after inactivity, active session not expired while in use, error response contains no internal data

---

RFC: THV-0038 — Session-scoped client lifecycle