fix(gemini,anthropic): base64 validation, escaped model paths, prompt-block mapping, unified unsupported-param semantics

testdata/anthropic/chat_logit_bias_empty_in.json

@@ -0,0 +1,5 @@
+{
+  "model": "claude-opus-5",
+  "messages": [{"role": "user", "content": "hi"}],
+  "logit_bias": {}
+}

testdata/anthropic/chat_logit_bias_empty_want.json

@@ -0,0 +1,5 @@
+{
+  "model": "claude-opus-5",
+  "messages": [{"role": "user", "content": "hi"}],
+  "max_tokens": 4096
+}

testdata/anthropic/chat_logprobs_false_in.json

@@ -0,0 +1,5 @@
+{
+  "model": "claude-opus-5",
+  "messages": [{"role": "user", "content": "hi"}],
+  "logprobs": false
+}

testdata/anthropic/chat_logprobs_false_want.json

@@ -0,0 +1,5 @@
+{
+  "model": "claude-opus-5",
+  "messages": [{"role": "user", "content": "hi"}],
+  "max_tokens": 4096
+}