fix: add jitter to retry backoff delays

[0xneobyte]

Problem

Retries use pure exponential backoff with fixed delays: 1s, 2s, 4s. When multiple clients encounter an error simultaneously (e.g. a brief backend outage), they all retry at exactly the same intervals. This creates a thundering herd — all clients hit the server at once on each retry wave, making recovery slower.

Proposed Behaviour

Add random jitter to retry delays to spread retries across a time window.

Current: delay = min(1000 * 2^attempt, 10000)
Fixed: delay = random(0, min(1000 * 2^attempt, 10000))

Full jitter is the most effective strategy for preventing thundering herd while keeping average latency reasonable.

Files to Modify

File	Change
src/brainus_ai/client.py	Add jitter to sleep duration in _make_request

Acceptance Criteria

• Retry delays are randomised within the exponential backoff window
• Maximum possible delay is unchanged (10s cap)
• Minimum possible delay is 0
• Retry count behaviour is unchanged
• Tests updated to account for non-deterministic delay