Consolidate fragmented memory system into a single foundation/memory/ package

## Problem

The memory system is spread across multiple locations with overlapping responsibilities:

| File | Purpose |
|------|---------|
| `foundation/memory.py` | `SessionMemory` (in-memory list) + `WorkspaceMemory` (file-based `.md`) |
| `src/platform_agent/memory.py` | Older/alternative memory implementation |
| `foundation/memory_access_guard.py` | Namespace-based access control |
| `foundation/hooks/memory_hook.py` | Memory lifecycle integration |
| `foundation/hooks/memory_extraction_hook.py` | STM→LTM extraction pipeline |
| `foundation/hooks/consolidation_hook.py` | Memory consolidation |

This fragmentation makes it difficult for contributors to understand the full memory architecture and for users to know which memory class to use.

### Additional issue: rough token estimation

`SessionMemory` estimates tokens with `~4 chars per token for English text`. This undercounts for code-heavy conversations (~3 chars/token) and overcounts for non-English text. Since the compaction threshold depends on this estimate, inaccuracy can cause premature or delayed compaction.

## Proposed Solution

Consolidate into a single `foundation/memory/` package:

```
foundation/memory/
├── __init__.py # Public API: SessionMemory, WorkspaceMemory, LongTermMemory
├── session.py # SessionMemory (in-memory conversation history)
├── workspace.py # WorkspaceMemory (file-based .md persistence)
├── long_term.py # AgentCore LTM integration (STM→LTM pipeline)
├── access_guard.py # MemoryAccessGuard (namespace security)
└── extraction.py # Memory extraction strategies
```

Move memory-related hooks to use the consolidated package. Deprecate `src/platform_agent/memory.py` with a shim pointing to the new location.

For token estimation, either:
- Use `tiktoken` for accurate counts
- Or expose `chars_per_token` as a configurable parameter (default 4, recommend 3 for code-heavy use cases)

## Acceptance Criteria

- [ ] All memory code consolidated under `foundation/memory/`
- [ ] `src/platform_agent/memory.py` deprecated with import shim
- [ ] Token estimation is configurable or uses a proper tokenizer
- [ ] Memory hooks import from the consolidated package
- [ ] All existing memory tests pass against the new structure
- [ ] ARCHITECTURE.md updated to reflect the new layout

Start by reading foundation/memory.py, src/platform_agent/memory.py, the memory hooks, and the existing memory tests to map current imports and behavior. Consolidate the listed components under foundation/memory/, preserve compatibility through the deprecated shim, address configurable or tokenizer-based token estimation, and update ARCHITECTURE.md; done means the acceptance criteria are met and all existing memory tests pass.

Written by the indexing model from the issue text.