Migrate Lingo memory ownership to Strata - #2

Summary

Migrates Lingo's bounded registry allocation to the shared ZekStack Strata memory policy and prepares Lingo v0.2.0.

Memory policy

• add Strata v0.1.2 as a dependency
• replace LingoConfig::preferPsram with Strata::MemoryPolicy
• preserve the existing default with memory.allocation = Strata::Placement::PreferExternal
• support Default, Internal, PreferExternal, and RequireExternal
• validate the complete memory policy during init()

Ownership

• route the registry through Strata::allocateArray / Strata::free
• remove src/internal/LingoMemory.h
• remove direct ESP-IDF heap capability usage and PSRAM allocation stubs
• preserve the one-allocation-at-init design; registration and lookup remain allocation-free

Diagnostics

• add LingoDiag
• report requested registry Strata::Placement separately from observed Strata::Region

Tests and CI

• replace ESP heap mocking with Strata generic-backend memory policy tests
• add source auditing preventing direct heap ownership from returning
• pin Strata v0.1.2 in CI
• keep the ESP32 / S3 / C3 / P4 PlatformIO matrix
• add matching Arduino CLI builds
• isolate PlatformIO and Arduino CLI toolchain state under the runner temporary directory
• remove multi-gigabyte ~/.platformio / ~/.arduino15 cache restores that exhausted self-hosted runner disks
• clean temporary toolchain state after embedded builds
• cancel superseded push/PR CI runs for the same branch so duplicate heavy matrices do not compete for runner storage

Release/docs

• bump library metadata to v0.2.0
• add v0.1.x migration guidance
• update README, configuration, API, memory, troubleshooting, and examples

Compatibility

The former default preferPsram = true maps to Strata::Placement::PreferExternal, so normal runtime placement behavior is preserved. preferPsram = false migrates to Strata::Placement::Internal.

CI failure follow-up

The observed embedded matrix failure was caused by runner disk exhaustion, not by a Lingo/Strata compile error. GitHub Actions was restoring a roughly 3.4 GB PlatformIO cache into self-hosted runners that already contained persistent PlatformIO/Arduino state, eventually reaching 0 MB free and failing with No space left on device. The CI now uses disposable per-job toolchain directories and avoids those oversized persistent cache restores.