perf: cut request/response copies and unblock both transports (0.5.0)

[abcsnowwolf]

• keep the parsed request document and view params into it instead of deep-copying params into a second JsonDocument
• serialize responses directly rather than assembling them in a wrapper document, removing another full-payload deep copy per reply
• answer tools/list from a body serialized once per tool-set change, invalidated by RegisterTool; the cache is unsynchronized, so document the single-task invariant it relies on
• add parseRequest(const char*) so neither transport builds a throwaway std::string of the whole request body
• add RegisterTool(Tool&&) to skip the schema-tree deep copy
• add MCP_OMIT_TEXT_WHEN_STRUCTURED to drop the duplicated text mirror of a structured result, built and asserted by env:native-omit-text
• offer a 517-byte preferred ATT MTU (BleServerConfig::preferredMtu); fragment count scales directly with the negotiated MTU, now measured by a transport test rather than asserted in a comment
• default BleServerConfig::txGapTicks to 0, replacing the unconditional 1-tick inter-fragment delay that cost ~430 ms per 8 KiB message
• answer a tools/call inline when it completes within MCP_HTTP_FAST_PATH_WAIT_MS instead of waiting for the ~500 ms connection poll that drives the deferred filler
• raise the RX baseline to 512 and only release growth past 4x baseline, ending the grow/shrink churn around the old 256-byte baseline
• make the BLE worker stack size configurable (MCP_BLE_WORKER_STACK_SIZE)
• keep the tool map on plain std::map<std::string, Tool>: std::less<> is C++14 and MCPServer.h compiles as part of the consumer sketch, which the ESP32 Arduino framework still builds as gnu++11
• match the real NimBLE 2.x setMTU signature (bool, not int) in the mock
• add scripts/coverage.sh plus env:native-cov and a CI gate at 80% line coverage (currently 89.0%)