fix(tongyi): use DashScope keepalive session pool

[Jinx2000]

Summary

Follow-up to #3815 and #3817.

DashScope confirmed that its process-wide shared synchronous session could
reuse silently stale pooled connections, causing requests to block until the
default 300-second read timeout. The underlying issue was fixed in
dashscope/dashscope-sdk-python#187 and released in
DashScope SDK 1.27.4.

Tongyi 0.2.19 already locks DashScope 1.27.4, but the embedding implementation
still creates a separate requests.Session for each invocation and passes it
explicitly to the SDK. This bypasses the SDK-managed shared session and its new
keepalive-enabled HTTP adapter.

This change removes explicit session creation and injection from:

• Text embedding
• Multimodal embedding
• Embedding credential validation

Embedding requests now use the shared session managed by DashScope 1.27.4.
Existing batching, token accounting, connection-error retries, timeout
handling, and HTTP 429 handling remain unchanged.

The LLM and rerank implementations are not changed.

Release Notes

Use DashScope's keepalive-enabled shared connection pool for Tongyi text and
multimodal embedding requests, reducing the risk of long stalls caused by stale
idle connections.

Change Type

• Documentation / non-plugin change
• Non-LLM plugin (tools, extensions, datasource, etc.)
• LLM plugin

Screenshots / Videos

This is a backend connection-management change and has no user-interface
changes.

Before	After
Embedding requests could intermittently take approximately 301 seconds after an idle connection became stale.	The approximately 301-second stall was not reproduced with the validation build using DashScope's keepalive-enabled shared session.

LLM Plugin Checklist

Areas affected by this change (check all that apply)

• Message flow (system messages, user ↔ assistant turn-taking)
• Tool interaction flow (multi-round usage, Agent App and Agent Node)
• Multimodal input (images, PDFs, audio, video, etc.)
• Multimodal output (images, audio, video, etc.)
• Structured output (JSON, XML, etc.)
• Token consumption metrics
• Other LLM functionality (text and multimodal embedding connection management)
• New models / model parameter fixes

Version

• Bumped top-level version in manifest.yaml from 0.2.19 to 0.2.20
• dify_plugin>=0.3.0,<0.6.0 is declared in pyproject.toml and locked in uv.lock

Testing

• Local deployment — Dify version: 1.13.3 (self-hosted local build)
• SaaS (cloud.dify.ai)

Validation performed:

• Packaged the plugin from a clean Git archive using the Dify Plugin CLI.
• Installed frozen dependencies under Python 3.12.
• Confirmed that uv.lock resolves dashscope==1.27.4.
• Passed the official serverless plugin installation validator.
• Ran the complete test suite from the unpacked .difypkg.
• Test result: 167 passed, 83 skipped.
• Deployed the 0.2.20-dev1 validation package to two self-hosted
  plugin_daemon nodes.
• Text embedding and knowledge retrieval remained responsive after more than
  two hours of idle time, without reproducing the approximately 301-second stall.